local commutative algebra and hochschild cohomology ......local commutative algebra and hochschild...
Post on 09-Jun-2020
1 Views
Preview:
TRANSCRIPT
Local Commutative Algebra and Hochschild Cohomology Through theLens of Koszul Duality
by
Benjamin Peter Briggs
A thesis submitted in conformity with the requirementsfor the degree of Doctor of Philosophy
Department of MathematicsUniversity of Toronto
c© Copyright 2018 by Benjamin Peter Briggs
Abstract
Local Commutative Algebra and Hochschild Cohomology Through the Lens of Koszul Duality
Benjamin Peter Briggs
Doctor of Philosophy
Department of Mathematics
University of Toronto
2018
This thesis splits into two halves, the connecting theme being Koszul duality. The first part concerns
local commutative algebra. Koszul duality here manifests in the homotopy Lie algebra. In the second
part, which is joint work with Vincent Gelinas, we study Hochschild cohomology and its characteristic
action on the derived category.
We begin by defining the homotopy Lie algebra π∗(φ) of a local homomorphism φ (or of a ring) in
terms of minimal models, slightly generalising a classical theorem of Avramov. Then, starting with work
of Felix and Halperin, we introduce a notion of Lusternik-Schnirelmann category for local homomor-
phisms (and rings). In fact, to φ we associate a sequence cat0(φ) ≥ cat1(φ) ≥ cat2(φ) ≥ · · · each cati(φ)
being either a natural number or infinity. We prove that these numbers characterise weakly regular, com-
plete intersection, and (generalised) Golod homomorphisms. We present examples which demonstrate
how they can uncover interesting information about a homomorphism. We give methods for computing
these numbers, and in particular prove a positive characteristic version of Felix and Halperin’s Mapping
Theorem.
A motivating interest in L.S. category is that finiteness of cat2(φ) implies the existence of certain
six-term exact sequences of homotopy Lie algebras, following classical work of Avramov. We introduce
a variation π∗(φ) of the homotopy Lie algebra which enjoys long exact sequences in all situations, and
construct a comparison π∗(φ) → π∗(φ) which is often an isomorphism. This has various consequences;
for instance, we use it to characterise quasi-complete intersection homomorphisms entirely in terms of
the homotopy Lie algebra.
In the second part of this thesis we introduce a notion of A∞ centre for minimal A∞ algebras.
If A is an augmented algebra over a field k we show that the image of the natural homomorphism
χk : HH∗(A,A) → Ext∗A(k, k) is exactly the A∞ centre of A, generalising a theorem of Buchweitz,
Green, Snashall and Solberg from the case of a Koszul algebra. This is deduced as a consequence of a
much wider enrichment of the entire characteristic action χ : HH∗(A,A)→ Z(D(A)). We give a number
of representation theoretic applications.
ii
Acknowledgements
Thank you Ragnar, I have learned so much from you. It is still so strange ending all this while you are
not here.
Thank you Vincent, I obviously could not have done this without you.
Thank you Ozgur, you are unbearably kind.
Thank you to my family, absolutely nothing is more valuable.
Thank you Leonid, Jeff and Vanya. I would have stayed forever, but everything has to end.
Thank you Srikanth Iyengar and Sarah Witherspoon for going absolutely beyond what was necessary in
helping me.
Thank you Jemima, you tirelessly look after us because we obviously can’t look after ourselves.
Thank you Joel Kamnitzer for helping us all without hesitation when Ragnar passed away.
Thank you, the rest of you. Thank you, thank you, thank you, thanks.
iii
Contents
1 Introduction 1
1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Notation and Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Background 12
2.1 Graded Categories and Differential Graded Categories . . . . . . . . . . . . . . . . . . . . 12
2.2 Twisting Cochains and the Bar Construction . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Koszul Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 A∞ Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5 Models for the Hochschild Cochain Complex . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6 Semi-Free Extensions and Acyclic closures . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.7 Cohen Factorisations and Minimal Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.8 Lie Algebras and Lie Coalgebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 Koszul Duality in Local Commutative Algebra 46
3.1 The Homotopy Lie algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 The L.S. Category of a Ring Homomorphism . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 The Mapping Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4 Functoriality of the Tate model for the diagonal . . . . . . . . . . . . . . . . . . . . . . . . 67
3.5 Some Long Exact Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4 The Characteristic Action of Hochschild Cohomology 86
4.1 Enriching the Characteristic Action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2 The A∞ centre . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.3 Examples and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Bibliography 103
iv
Chapter 1
Introduction
This thesis splits roughly into two halves, the connecting theme being Koszul duality. The first part
concerns local commutative algebra, which is broadly the study of singularities. Koszul duality here man-
ifests in the homotopy Lie algebra, whose use in commutative algebra has been pioneered by Avramov.
The main character in the second part of this thesis is Hochschild cohomology, and its characteristic
action on the derived category. Structures known as A∞ algebras arise naturally through Koszul duality,
and we investigate how these can be used to understand and constrain the characteristic action. This
part of the thesis is joint work with Vincent Gelinas, and our contribution here should be considered
equal. It is adapted from [34].
Koszul duality is a wide reaching and mysterious phenomenon. In topology it connects a group
with its classifying space. In rational homotopy theory it explains the relationship between Sullivan’s
commutative model and Quillen’s Lie algebra model for a space. In algebra it yields unexpected and
useful derived equivalences. Rich examples and other Koszul duality like phenomena abound throughout
algebra and geometry.
In its simplest algebraic incarnation Koszul duality concerns the relationship between an associative
algebra A augmented over a field k, and the Yoneda algebra E = Ext∗A(k, k). When A is a Koszul
algebra (see section 2.3) this relationship is particularly striking. Perhaps the most beautiful aspect
is the symmetry: this hypothesis alone means that E is also Koszul, and that there is a canonical
isomorphism A ∼= Ext∗E(k, k). We will also see that the the representation theory of A is intimately
connected to that of E. We say that A and E are Koszul dual to each other. Aside from the rich
structure arising from this assumption, it turns out that Koszul algebras are extremely common.
A fundamental example is the Koszul duality between the polynomial algebra S = k[x0, ..., xn] and
the exterior algebra Λ =∧k(y0, ..., yn). It is this example which gives rise to the famous Bernstein-
Gel’fand-Gel’fand correspondence [30], which describes an equivalence between bounded complexes of
coherent sheaves on projective space Pnk and finitely generated Λ modules without free summands.
On the other hand, local commutative algebra concerns singularities. It is by now very well un-
derstood that one can use homological methods to unravel the character of a singularity. A funda-
mental illustration is that the violence of a singularity is reflected in the growth of the Betti numbers
1
Chapter 1. Introduction 2
βiR(k) = dimkExtiR(k, k), where R is the local ring of functions on the singularity and k is its residue
field.
These ideas go all the way back to Hilbert’s syzygy theorem, which says that if M is a finitely
generated module over a polynomial algebra S = k[x1, ..., xn] on some field k, then the Betti numbers
βiS(M) = dimkExtiS(M,k) vanish as soon as i > n. In modern language, we say that S has global
dimension n. In the mid 1950s Auslander, Buchbaum and Serre established a converse: a local ring R
has finite global dimension if and only if it is regular, that is, if it is the ring of functions at some smooth
point. By these means a number of long standing problems in commutative algebra fell under the force
of homological algebra.
The importance of the Betti numbers βiR(k) = dimkExtiR(k, k) suggests that Koszul duality is already
at play in local commutative algebra. In fact, the structure of Ext∗R(k, k) has been an important focal
point in this area for decades. One highlight is the cohomology operators of Gulliksen and Eisenbud. If R
is a complete intersection of codimension c, these operators χ1, ..., χc are central elements in Ext2R(k, k)
which act on the cohomology of any R module. They have turned out to be extraordinarily useful. In
short, they control the stable behaviour of modules over R. A good place to start is [15], where the
geometric nature of the cohomology operators is exploited.
Miraculously, from the hypothesis that R is a commutative it follows that the Yoneda algebra
Ext∗R(k, k) is the universal enveloping algebra of a certain graded Lie algebra. This fact is the cul-
mination of work of Milnor-Moore, Lofwall, Sjodin, Levin and Schoeller (see section 3.1). This object
bears a close and mysterious resemblance to the Whitehead Lie algebra π∗(ΩX) associated to a simply
connected based space X. As such it is known as the homotopy Lie algebra of R, and we denote it
π∗(R). It will be a central character in this thesis.
In the early 1980s Avramov and Roos began to uncover deep connections between local commutative
algebra and rational homotopy theory, centring around the homotopy Lie algebra. This has led to a
great deal of collaboration and progress in both areas, some of which is covered in the surveys [9] and
[6]. Much of what is now known about π∗(R) was inspired by the work of rational homotopy theorists,
although it is often much more difficult to establish a theorem in commutative algebra than its cousin in
rational homotopy theory (or vice versa). For example, the famous dichotomy theorem says that either
R is a complete intersection, in which case π≥3(R) = 0, or the sequence of numbers εi(R) = dimkπi(R)
must grow exponentially. This was proven by Avramov in [13] almost two decades after the dichotomy
theorem for finite CW complexes was established by Felix, Halperin and Thomas in [45].
But the homotopy Lie algebra has been important regardless of its connection to topology. For
instance, understanding the numbers εi(R) can lead to fine control of the Poincare series of modules over
R. As another example, when R is a complete intersection, the cohomology operators themselves come
from π2(R). Likely the most famous application of the homotopy Lie algebra is Avramov’s use of it in
[13] to solve a long standing conjecture of Quillen. In his groundbreaking development of the simplicially
defined cotangent (or Andre-Quillen) homology, Quillen posed the following question: if φ : R → S is
a ring homomorphism which is essentially of finite type such that S has finite flat dimension as an
R module, and such that the cotangent complex Lφ has finite projective dimension, then must φ be
a locally-complete intersection? The homotopy Lie algebra is usually much easier to compute than
cotangent homology. Avramov was able to pass some of this information through a spectral sequence
from the dg to the simplicial world. Via this channel the dichotomy theorem for the homotopy Lie
algebra was used to settle Quillen’s quastion.
Chapter 1. Introduction 3
Long before Quillen’s foundational work in rational homotopy theory, the connection between com-
mutative algebras and Lie algebras was exploited by Chevalley and Eilenberg. The famous Chevalley-
Eilenberg cochain complex C∗(L;Q), which computes the Lie algebra cohomology of L, is an example
of a minimal model when L is nilpotent. We will say what this means shortly (in the dual, homological
context), and much more detail will be given in section 2.6.
After the work of Quillen, Sullivan’s use of minimal models in rational homotopy theory was an
important break though, both theoretically and computationally. It was once again Avramov who
imported minimal models into local commutative algebra. They will be the fundamental objects of
study in chapter 3. I’ll explain now how they can be used to compute the homotopy Lie algebra, and
then go over some highlights from chapter 3.
Let X be a graded set of variables in strictly positive degrees. We denote by Q[X] the free strictly
graded commutative Q algebra on X. This is the exterior algebra over Q on odd elements of X tensored
with the polynomial algebra on even elements of X. A dg algebra over Q will be called semi-free if it
has the form Q[X] after forgetting its differential. These are non-linear analogues of free resolutions for
modules. In particular, they can be used to resolve local rings.
For us, a minimal model is a semi-free dg algebra A = Q[X] over a regular local ring Q which satisfies
the minimality condition ∂(mA) ⊆ m2A, where mA is the unique maximal, graded ideal of A. Intuitively,
these dg algebras are as close to being free as possible, because regular local rings behave homologically
as if they were free.
If R is a local ring then by the Cohen structure theorem there is a regular local ring Q with the
same embedding dimension of R, and a surjective homomorphism Q → R onto the completion of R.
It is possible to inductively construct a minimal model A = Q[X] and a surjective quasi-isomorphism
A'−→ R of Q algebras.
Note that the indecomposables mA/m2A of A naturally form graded vector space over k = A/mA.
Theorem 1 (theorem 15). Suppose that A = Q[X] is a minimal model resolving R as above. The shifted
and dualised space of indecompables π∗(A) = (ΣmA/m2A)∨ canonically has the structure of a graded Lie
algebra over k, and there is a natural isomorphism Uπ∗(A) ∼= Ext∗R(k, k).
This extends a classical theorem of Avramov [9, theorem 4.2] which deals with π≥2(A). It follows
that π∗(A) is independent of the construction of A and thus we denote it π∗(R). This is the homotopy
Lie algebra of R.
Using Avramov, Foxby and Herzog’s theory of Cohen factorisations one can also associate a minimal
model A to a local homomorphism φ : R→ S. In this way one obtains an analogous graded Lie algebra
π∗(A) = π∗(φ), which we call the homotopy Lie algebra of φ.
In rational homotopy it is necessary to impose conditions on a space to obtain a well-behaved homo-
topy Lie algebra. Indeed, Quillen [105] has shown that any finite type graded Lie algebra over Q occurs
as π∗(ΩX) for some rational space X, which means no structure theory is possible in general. After
the work of Felix and Halperin (and a number of other authors, see the beginning of section 3.2) we
know that π∗(ΩX) enjoys some surprising rigidity properties when X has finite Lusternik-Schnirelmann
category, which simply means that X admits a finite cover by open sets, each of which is contractible
in X. These authors also characterised the L.S. category of a rational space algebraically in terms of
minimal models [43].
Starting with the work of Felix and Halperin we discuss a notion of L.S. category for local rings and
Chapter 1. Introduction 4
local homomorphisms. In fact, to a local ring R and a local homomorphism φ we associate sequences
cat0(R), cat1(R), cat2(R), · · · and cat0(φ), cat1(φ), cat2(φ), · · ·
Each cati is either a natural number or infinity, and we’ll see that the sequences are in fact weakly
decreasing (this is a consequence of the Mapping Theorem). After establishing well-definedness we prove
that the vanishing of these numbers characterises regularity and the complete intersection property. We
also show that these invariants characterise the Golod property. We also give several examples which
demonstrate how the behaviour of the L.S. category can uncover interesting information about a local
ring of homomorphism. This part of the thesis especially is just the beginning of what will hopefully
become a more complete development of L.S. category in local commutative algebra.
Our main interest in L.S. category is that its finiteness implies that the homotopy Lie algebra π∗(φ)
has good structural properties, after the important work of Felix, Halperin, Jacobsson, Lofwall and
Thomas [44] (see theorem 18). Using classical work of Avramov one can also build certain six term
exact sequences of homotopy Lie algebras under the hypothesis that cat2(φ) is finite. This is the subject
of section 3.5.
In [43] Felix and Halperin prove the Mapping Theorem, an important tool for calculating L.S. category
in rational homotopy theory. The corresponding fact in local commutative algebra has been missing for
some time. It turns out their result generalises verbatim to the situation of positive residual characteristic
(even over a regular local base ring), we give a proof in section 3.3.
Theorem (The Mapping Theorem 20). If A → B is a surjective homomorphism of minimal models
then catA ≥ catB.
The proof of the classical mapping theorem uses good properties of the category of dg commutative
algebras over Q which are not available in our context. Thus, the proof we give is necessarily more
delicate.
This version of the Mapping theorem generalises theorem 1.2 of [19], among other things. More
importantly the theorem can be used to construct new ring homomorphisms of finite L.S. category. In
this situation the deep results of [44] then apply, strongly constraining the algebraic structure of π∗(φ).
The proof also applies in the dual, cohomological situation of rational homotopy theory, but now over a
field of any characteristic. This has potential applications to homotopy theory at large primes, after the
work of Anick [4] (but that direction will not be pursued in this thesis).
Jumping ahead to section 3.5, we investigate long exact sequences for the homotopy Lie algebra, and
for variations on it. An important construction we make in that section is the following.
Theorem (proposition 12 and theorem30). To a local homomorphism φ : (R, k) → (S, l) one can
associate a graded vector space π∗(φ) with the following properties. If (Q, u)ψ−→ (R, k)
φ−→ (S, l) is a
sequence of local homomorphisms there is a long exact sequence
· · · → πi−1(ψ)⊗k l→ πi(φ)→ πi(φψ)→ πi(ψ)⊗k l→ πi+1(φ)→ · · ·
There is a natural comparison map π∗(φ)→ π∗(φ) which is an isomorphism whenever cat2(φ) <∞.
One of the reasons for introducing π∗(φ) is that it also recovers some other known long exact sequences
which do not involve π∗(φ), but another variation defined using divided powers instead of free variables.
Details on this are given in section 3.5.
Chapter 1. Introduction 5
We show that vanishing of π∗(φ) characterises quasi-complete intersection homomorphisms and de-
duce the following characterisation as a corollary.
Theorem (theorem 33). Let φ : (R, k) → (S, l) be a local homomorphism. Then φ is quasi-complete
intersection if and only if πi(R)⊗k l→ πi(S) is an isomorphism for i ≥ 3.
This establishes the converse to a theorem of Avramov, Henriques and Sega [18, theorem 5.3]. It
means that the quasi-complete intersection property can in fact be detected by the homotopy Lie algebra.
One of the key technical tools is the Tate model for the diagonal, which associates to a minimal model
A = Q[X] a minimal dg module of A over A ⊗Q A. This minimal resolution turns out to have a dg
algebra structure, and is in fact free as a divided power algebra over A⊗QA. This classical construction
is recalled in section 2.6, where a few applications are discussed.
The Tate model for the diagonal involves both free and divided power variables. Building homomor-
phisms between such dg algebras is often a serious technical problem. In section 3.4 we deal with this
problem by establishing that the Tate model for the diagonal is in a certain sense functorial. This proves
to be very useful, and some applications are given in that section.
Theorem (theorem 31). Let φ : (R, k)→ (S, l) be a local homomorphism. If φi : πi(S)→ πi(R)⊗k l is
surjective for all even i then there are six term exact sequences
0 π2i−1(φ) π2i−1(S) π2i−1(R)⊗k l
π2i(φ) π2i(S) π2i(R)⊗k l 0.δ
After [8], the homomorphism φ is called small if φ∗ : π∗(S) → π∗(R) ⊗k l is surjective. It is known
that these for these homomorphisms one obtains three term exact sequences as above (with δ = 0). It
is striking that also surjectivity in even degrees is equivalent to the presence of these six term exact
sequences, even though there is a priori no long exact sequence at all. As a consequence of this theorem,
if (Q, u)ψ−→ (R, k)
φ−→ (S, l) are local homomorphisms such that both cat2(ψ) and cat2(φ) are finite, we
prove that there is a long exact sequence
· · · → πi−1(ψ)⊗k l→ πi(φ)→ πi(φψ)→ πi(ψ)⊗k l→ πi+1(φ)→ · · ·
of homotopy Lie algebras (see corollary 6).
In chapter 4 the theme changes dramatically, our central interest from this point on in the char-
acteristic action of the Hochschild cohomology HH∗(A,A) on the derived category an algebra A. The
underlying theme remains Koszul duality. This part of the thesis is adapted from [34], and as such its
content is joint work with Vincent Gelinas.
At this point it is wise to remember that Koszul duality originates independently from topology.
Here it concerns the relation between a based space X and its loop space Y = ΩX. Keeping track of
certain homotopical data, one can recover X, up to homotopy, as a classifying space BY .
In algebra, B and Ω become the bar and cobar functors. They are indispensable tools for us below.
The ideas surrounding them are extremely wide reaching, and Koszul duality manifests in many places
through some kind of bar-cobar formalism. Many of these ideas were first codified in [55].
Chapter 1. Introduction 6
It is through these functors that one arrives at the idea of an A∞ algebra, a flexible generalisation of
associative algebras defined by Stasheff with topological motivation. Roughly, an A∞ algebra possesses
a sequence of n-ary operations for n ≥ 2, rather than only a binary multiplication, that together describe
a structure which is associative up to homotopy. These structures provide us with the framework we
need to interpret our results on the characteristic action, as we will explain now.
The Hochschild cohomology HH∗(A,A) acts centrally on the derived category of A through the
characteristic morphism χM : HH∗(A,A) → Z(Ext∗A(M,M)) for each dg module M . This action has
a great many uses, controlling much of the homological behaviour of M . To name a few directions, it
has been studied in connection with the cohomology operators of Gulliksen [14]; the theory of support
varieties [112]; with obstructions to deforming modules; it connects to geometry via the Atiyah-Chern
character [37]; and in rational homotopy theory it shows up as an Umkehr map to the loop homology
algebra [49].
Despite this, the image of χM : HH∗(A,A) → Ext∗A(M,M) is not well understood beyond the fact
that im(χM ) ⊆ Z(Ext∗A(M,M)). We investigate how the image of χM is constrained by the natural A∞
algebra structure on Ext∗A(M,M).
To this end we introduced a notion of A∞ centre for a minimal A∞ algebra A. In short, this can
be defined by setting Z∞(A) to be the image of the canonical projection Π : HH∗(A,A)→ A. It is also
possible to define Z∞(A) using “higher commutators” for the higher products of A.
This is perhaps best viewed as a computational tool. For example, one whether certain obstructions
vanish (see remark 20), or establishing the finite generation conditions which are vital in the theory
of support varieties (see for example [112]). Specifically, when is HH∗(A,A) Noetherian, and when is
Ext∗A(k, k) module-finite over the image of χk?
Once again this goes back to the work of Quillen, who introduced support theory to the modular rep-
resentation theory of finite groups. Here it is vital to know that the cohomology algebra of a finite group
is finitely generated, as established by Venkov and Evens. More generally, Snashall and Solberg conjec-
tured in [112] that for every finite dimension k algebra A the Hochschild cohomology HH∗(A,A)/(nil)
modulo nilpotents should be a finite generated k algebra. This is know for many classes of algebras (see
op. cit.), but a counter-example has since been produced by Xu [123]. Aside from computing the image
of χk, in some cases our methods are helpful in describing the algebra HH∗(A,A)/(nil) (see corollary
12).
Now we come to the original motivation for this work. Buchweitz, Green, Snashall and Solberg [38]
have shown that if A is a Koszul algebra then the image of χk : HH∗(A,A)→ Ext∗A(k, k) is the graded
centre Z(Ext∗A(k, k)). Their theorem is generalised, and given a more conceptual proof, in the following.
The connectedness assumption will be explained in section 2.3.
Theorem (theorem 37). If A is a strongly connected augmented dg algebra then the image of χ :
HH∗(A,A)→ Ext∗A(k, k) is exactly Z∞(Ext∗A(k, k)).
Since Koszulity of A can be characterized by formality of RHomA(k, k), and since the A∞ centre
coincides with the graded centre for honest graded algebras, the theorem of Buchweitz, Green, Snashall
and Solberg is precisely the special case when A is Koszul.
Let us emphasize the computational nature of our definitions. The A∞-centre is concretely definable
in terms of certain higher commutators. Moreover, the natural A∞ structure on Ext∗A(k, k) be uncovered
algorithmically, and in many cases very explicitly. This is so for the d-Koszul algebras of Berger [27],
after the work of He and Lu [66] (see section 4.3), and also for monomial quiver algebras, after Tamaroff
Chapter 1. Introduction 7
[116]. Knowing this structure, theorem 37 can often be used to compute the image of χk completely and
explicitly by hand.
Theorem 37 can be generalised to arbitrary dg modules, rather than the augmentation module.
Theorem (corollary 8). Let M be a right dg module over A, and take Ext∗A(M,M) to be a minimal A∞
model for RHomA(M,M). For every dg module M , the characteristic morphism χM : HH∗(A,A) →ExtA(M,M) lands in the A∞ centre of Ext∗A(M,M). If M is homologically balanced as an REndA(M)−A bimodule then im(χM ) = Z∞(Ext∗A(M,M)).
The terminology homologically balanced is explained in section 2.1. We recover theorem 37 as the
special case that M = k because k is homologically balanced whenever A is strongly connected (see
section 2.3). Corollary 8 is deduced from the following theorem.
Theorem. Let A be a dg algebra augmented over a field k, and let M be a right dg module over A.
Denote the derived endomorphism algebra of M by R = REndA(M,M). The characteristic morphism
χM lifts canonically into the Hochschild cohomology of R:
HH∗(A,A) HH∗(R,R)
Z(Ext∗A(M,M)).
χM
χM ΠR
Moreover, if M is homologically balanced then χM is an isomorphism.
This in turn is a consequence of a larger enrichment χ for the entire characteristic action at once
(theorems 34 and 35). This comes from a detailed investigation of various models for the characteristic
action of HH∗(A,A) on D(A). A consequence is that this action is actually independent of A: it depends
only on the enhanced derived category Ddg(A) up to quasi-equivalence.
In the final section of this thesis we compute a few examples showing theorem 37 in action. This
section 4.3 also contains a few applications to specific classes of algebras. We also return to the question
of finite generation of HH∗(A,A)/(nil). The d-Koszul algebras appearing in the theorem below are a
natural non-quadratic generalisation of Koszul algebras, they will be defined in section 37.
Theorem (theorem 40 and corollary 12). Assume that A is d-Koszul with nilpotent augmentation ideal,
then in even cohomological degrees χ surjects onto the graded centre
χ : HHeven(A,A) Z(ExtevenA (k, k))
and this induces an isomorphism of graded commutative algebras
HH∗(A,A)/(nil) ∼= Z(ExtA(k,k))/(nil).
Finally, let us point our another application taken from section 4.2. The proof uses A∞ structures
and the ideas behind the A∞ centre in an essential way, but we can state it purely in dg language.
Theorem (corollary 10). Assume k has characteristic zero and let L be a dg Lie algebra over k. Then
Π : HH∗(UL,UL)→ H∗(UL) is surjective if and only if L is formal and quasi-isomorphic to an abelian
Lie algebra.
Chapter 1. Introduction 8
In the literature these dg Lie algebras have been called quasi-abelian. Understanding when this
condition holds is important because it means the associated deformation problem is extremely simple,
and in particular unobstructed.
1.1 Overview
The first background section 2.1 is a rapid account of some standard definitions in homological algebra,
the most important being the centre of a graded category. It could be skimmed or skipped entirely.
After this, sections 2.2, 2.3, 2.4, and 2.5 contain a more leisurely introduction to the basic concepts
surrounding Koszul duality. They should be read in order. Since the content here is relatively standard
an expert may also wish to skim or skip these sections.
Section 2.2 is an introduction to twisting cochains and the bar construction which will be relevant
throughout this thesis. Section 2.3 contains some highlights from the classical theory of Koszul duality.
It is needed for chapter 4 only, but it will likely provide helpful context for both parts of this thesis.
Then in sections 2.4 and 2.5 we cover the background on A∞ algebras and Hochschild cohomology which
will be needed in chapter 4.
The material in these background sections is classical and none of it should be considered original,
except possibly parts of the presentation of twisted tensor products and of the Hochschild cochain
complex, which are adapted from [34, sections 2.1 and 2.3].
The background sections 2.6, 2.7 and 2.8 concern some important aspects of the homological algebra
of local rings, setting the scene for chapter 3. These sections should also be read in order. Much of the
content here is non-standard, so these sections should be read before chapter 3.
In section 2.6 we introduce semi-free extensions and semi-free divided power extensions. The back-
ground on divided powers is apparently more general than any easily accessible reference, but this
generality is largely superficial. This section also contains a quite general form of the famous Tate model
for the diagonal, followed by a few applications.
Section 2.7 begins with an introduction to the Cohen Factorisations of [16]. After this the important
class of minimal models is introduced. This section also contains a discussion of Golod homomorphisms.
Finally, 2.8 contains background on graded Lie algebras, leading directly on to section 3.1 on the
homotopy Lie algebra. The content is standard, except that theorem 13 (and its corollary) most likely
does not appear in the literature in this generality.
To summarise, the reader should be aware of the following main dependencies.
dgcategories
2.1
twistingcochains
2.2
Koszulduality
2.3
A∞algebras
2.4
Hochschildcohomology
2.5
semi-freeextensions
2.6
minimalmodels
2.7
Liealgebras
2.8
Hochschild cohomologychapter 4
commutative Koszul dualitychapter 3
In chapter 3 we investigate several related topics in local commutative algebra.
Chapter 1. Introduction 9
Section 3.1 leads directly on from the background section 2.8. It contains the fundamental construc-
tions and results describing the homotopy Lie algebra of a local ring or local homomorphism.
In section 3.2 we introduce and study the LS category for a local ring or ring homomorphism. This
is really the beginning of a full development. We establish a series of interesting properties and give
several examples. The proof of well-definedness is technical and delayed until the end of the section.
One of the most important ways to calculate LS category is the mapping theorem, which is established
in section 3.3.
A key tool throughout this thesis is the Tate model for the diagonal. The purpose of section 3.4 is
the proof a technical result which shows that the Tate model for the diagonal can in some sense be made
functorial. We use this result a number of times in the thesis. In particular, we explain in this section
how it can be used to give a simpler proof of a result of Avramov and Iyengar.
In the last section 3.5 of chapter 3 we discuss in detail some of the long exact exact sequences which
can be built from the homotopy Lie algebra, and from variations on it.
Chapter 4 concerns the characteristic action of Hochschild cohomology HH∗(A,A) on the derived
category D(A), and how this action interacts with Koszul duality for A.
The core of this chapter is section 4.1. Here we use a convolution enhancement for the derived
category, and the yoga of twisting cochains, to explain how the characteristic action can be enriched in
a strong sense.
Section 4.2 is about interpreting the previous section through the lens of A∞ structures. Here we
introduce and begin to study a notion of A∞ centre. As an application, at the end of the section we
give a criterion for when a dg Lie algebra is quasi-abelian.
Finally, the philosophy of this chapter is applied to a few examples of d-Koszul algebras in section
4.3. This section also summarises a few quick applications of the preceding sections.
1.2 Notation and Conventions
We briefly indicate in this section some of the notation and basic objects that will be used throughout
the rest of this thesis. These definitions are far from complete, they are here largely to make clear which
set of conventions will be adhered to below. We also sneak in a couple of fundamental lemmas.
A graded object in a category C is simply an indexed sequence X = Xii∈Z of objects of C. By
default indices are written as subscripts, and we say X is homologically graded. When an object is
naturally cohomologically graded we may use superscripts according to the convention Xi = X−i. A
graded object X has a suspension ΣX defined by the equalities (ΣX)i = Xi−1. When the objects Xi have
an underlying set, elements x ∈ Xi are said to have degree i and we write |x| = i. The corresponding
element of (ΣX)i+1 is written Σx.
If k is a commutative ring, then a complex of k modules is a graded k module M equipped with a
morphism ∂ : M → ΣM such that ∂2 = 0. The suspension ΣM is itself a complex with ∂(Σx) = −Σ∂(x).
We use the notation supM = supi : Hi(M) 6= 0. The cone on a chain map f : M → N is the complex
cone(f) = N + ΣM with the differential ∂(n + Σm) = ∂(n) − Σ∂(m) + f(m). A quasi-isomorphism
f : M → N is a chain map such that the induced map in homology H∗(f) : H∗(M) → H∗(N) is an
isomorphism. Throughout, the symbol ' will be used to indicate a quasi-isomorphism (in a slight abuse
of notation, we do not worry about the fact that this symbol is undirected). The following lemma
Chapter 1. Introduction 10
will be extremely useful for detecting quasi-isomorphisms. It can be proven by hand, or using the
Eilenberg-Moore comparison theorem [120, theorem 5.5.11].
Lemma 1. Let M and N be complexes with complete, exhaustive filtrations, and let f : M → N be
a map of filtered complexes. If the map gr(f) : gr(M) → gr(N) of associated graded complexes is a
quasi-isomorphism (or isomorphism, respectively) then so is f .
An (associative) dg algebra is a complex A of k modules with a product m : A ⊗ A → A and a
unit k → A which are both chain maps, and which satisfy the usual associativity and unitality axioms.
Dually, an (associative) dg coalgebra is a complex C of k modules with a coproduct ∆ : C → C ⊗ Cand a counit C → k which satisfy the dual associativity and unitality axioms. Details can be found for
instance in [90, chapter 1]. A right dg comodule over C is a complex M with a coaction ∆ : M →M⊗Cwhich is associative and unital. The situation for left comodules and bicomodules is analogous.
Dual to the usual definition for algebras, C is augmented if it is equipped with a homomorphism
k → C of coalgebras which splits the counit C → k. From an augmentation C inherits an increasing
filtration defined as follows. Write C = ker(C → k), since C is augmented there is an induced nonunital
coproduct ∆ : C → C ⊗ C. Using associativity to iterate yields the n-fold coproduct ∆(n)
: C → C⊗n
.
With this we define C(n) = k + ker(∆(n)
) for n ≥ 1. Then k = C(1) ⊆ C(2) ⊆ · · · ⊆ C is called the
primitive filtration and we say that C is cocomplete if⋃C(n) = C. This is the case for most coalgebras
which come up in practice, and it will be a crucial assumption below.
An important example is the tensor coalgebra Tco(V ) =⊕
w≥0 V⊗w on a graded k module V .
We use the notation πw : Tco(V ) → V ⊗w for the natural projections. The coproduct is given by
∆(v1 ⊗ · · · ⊗ vw) =∑wi=0(v1 ⊗ · · · ⊗ vi) ⊗ (vi+1 ⊗ · · · ⊗ vw). With the projection π1 it is cofree among
cocomplete coalgebras equipped with a linear map to V .
Let M be a bicomodule over a coalgebra C, with right and left coactions ∆r : M → M ⊗ C and
∆l : M → C ⊗M . The compatibility condition asserts that (∆l ⊗ 1)∆r = (1 ⊗∆r)∆l and we denote
this by ∆(3). A k linear map q : M → C is by definition a coderivation if ∆q = (q ⊗ 1)∆r + (1⊗ q)∆l.
The space of coderivations from M to C is denoted coDer(M,C). We will need the following
Lemma 2 (Quillen [107, part II section 3.1]). Let M be a bicomodule over Tco(V ). The corestriction
coDer(M,Tco(V )) → Hom(M,V ), q 7→ π1q is an isomorphism. Its inverse lifts α : M → V to the
coderivation
q =∑i,j≥0
(πi ⊗ α⊗ πj)∆(3).
Now we list some of our basic commutative algebra conventions.
A local ring is by definition a Noetherian commutative ring with exactly one maximal ideal. We will
often say that (R,m, k) is a local ring to mean R is a local ring with maximal ideal m and residue field
k. This shorthand applies as well to (R,m) or (R, k). At the same time, to avoid coming up with names
for maximal ideals we will often automatically denote by mR the maximal ideal of a local ring R.
If (R, k) is a local ring the Poincare series of a finitely generated R module M is by defintion
PRM (t) =∑i≥0 dimk TorRi (M,k)ti.
If a base ring is not specified then by dg algebra we mean dg algebra over Z (that is, a dg ring).
In chapter 3 most of the dg algebras we deal with will be strictly commutative. This means any two
Chapter 1. Introduction 11
elements satisfy xy = (−1)|x||y|yx and furthermore x2 = 0 whenever |x| is odd1. For us, a local dg
algebra A is by definition a strictly commutative dg algebra concentrated in nonnegative degrees such
that A0 is local and each Ai is finite as a module over A0. In particular, A has exactly one maximal
ideal. As for local rings, we use the notation mA by default for this ideal.
A homomorphism A → B of local dg algebras is called minimal if ∂(mB) ⊆ mAB + m2B . The dg
algebra A will be called absolutely minimal if ∂(mA) ⊆ m2A.
1In characteristic 2 this kind of commutativity is not even controlled by an operad (let alone a Σ-split operad), makingits meaning homotopically unclear. Rather, this condition is a direct translation of the rigid structures which one workswith in rational homotopy theory.
Chapter 2
Background
2.1 Graded Categories and Differential Graded Categories
In this short section we list some definitions concerning graded categories and dg categories over a
commutative ring k. In particular, we make explicit how some standard definitions for dg algebras and
dg modules generalise to the case of many objects. This will be somewhat terse; the coming sections
should (hopefully) be more readable.
Undecorated tensor product and Hom’s are all taken over k.
A graded k category C consists of a collection of objects and for any two objects x, y a graded k
module HomC(x, y), which will often be denoted yCx. We should also have bilinear composition maps
xCy ⊗ yCz → xCz and units k → xCx which satisfy the usual associativity and unitality conditions.
The underlying ungraded k category C0 has the same objects as C and HomC0(x, y) = HomC(x, y)0.
If there is an equivalence Σ : C0 → C0 and a natural isomorphism HomC(x, Σy) ∼= ΣHomC(x, y) then we
say that C has an internal suspension. In this way a graded category is essentially the same thing as a
k category with a distinguished auto-equivalence. In particular, every triangulated category is naturally
a graded category.
By definition the graded centre Z(C) of C is the graded k module of graded natural transformations
1C → 1C . This means
Zn(C) = families ξx ∈ HomC(x, x)nx∈C : fξx = (−1)nmξyf for all f ∈ HomC(x, y)m .
If C has a single object then C is simply a graded k algebra, and Z(C) is the usual graded centre of C.
On the other hand, if C has an internal suspension Σ then Zn(C) can be identified with the set of all
natural transformations ξ : 1C0→ Σ
n which satisfy ξΣ = (−1)nΣξ (that is, ξΣxΣ = (−1)nΣξx : x → Σn+1x
for all objects x). Under composition Z(C) becomes a graded commutative k algebra (not necessarily
strictly graded commutative). Krause and Ye discuss graded centres of triangulated categories in [83].
For any object x of C there is a restriction homomorphism Z(C)→ Z(EndC(x)). More generally one
can restrict along any full subcategory of C (in [79] Keller upgrades this with the observation that one
can also do this for the Hochschild cochain complex of any dg category, see section 2.5).
A dg category C over k consists of a collection of objects and a complex yCx = HomC(x, y) of k
modules for any objects x, y. Again C should come with composition maps xCy ⊗ yCz → xCz and
12
Chapter 2. Background 13
units k → xCx, both of which are chain maps, and which satisfy the usual associativity and unitality
conditions.
Any dg category has an underlying graded category by forgetting the differential. More interest-
ingly, taking homology produces a graded category H∗(C) with the same objects and HomH∗(C)(x, y) =
H∗(HomC(x, y)). The underlying ungraded category H0(C) is usually called the homotopy category of
C.
A functor F : C → D between dg categories is a functor of the underlying ungraded categories which
additionally respects differentials, meaning that HomC(x, y) → HomD(F (x), F (y)) is a chain map for
any two objects x and y of C. If the induced functor H∗(F ) : H∗(C)→ H∗(D) is an equivalence then F
is called a quasi-equivalence.
A right module M over a dg small category C is a complex Mx for every object x of C together
with a chain map Mx ⊗ xCy → My for any two objects x and y, and this action should be associative
and unital in the obvious sense. Similarly one can define left modules and bimodules. C is canonically
a bimodule over itself, and for any object x we have the representable right module xC (or left module
Cx). The representable modules play the role of free modules. Given a right module M and a left
module N over C, one can form the tensor product M ⊗C N , which is the quotient of⊕
x∈CMx ⊗ xN
by the identification m⊗ fn = mf ⊗ n for m ∈Mx and f ∈ xCy and n ∈ yN .
The preceding definitions can be put into a formal framework following [80]. A dg quiver Q consists
of a set of objects and a complex yQx for each pair of objects. The collection of dg quivers over a fixed
set of objects S forms a monoidal dg category by setting
Hom(Q,Q′) =∏x,y
Hom(yQx, yQ′x) y(Q⊗Q′)x =
⊕z
yQz ⊗ zQ′x.
The unit is the quiver kS with x(kS)x = k · 1x and y(kS)x = 0 when x 6= y. A dg category with a set
of objects S is then essentially the same thing as a monoid with respect to this structure, and modules
over dg categories are essentially modules over the corresponding monoids.
A dg category C on a set S of objects is called augmented if it is equipped with a morphism C → kS
splitting the unit kS → C.
One reason for introducing this framework is that it reveals the correct definition of dg cocategory:
these should be dg quivers with a decomposition C → C ⊗ C and a counit C → kS satisfying the usual
comonoid axioms. Cocategories can be used to build bar constructions for augmented dg categories.
A fundamental example of a dg category is the enhancement of the derived category of an algebra.
To describe this we need some background. These dg categories are never augmented.
Let A be a dg algebra (or small dg category) over k. A right module M over A is called semi-free
if it admits an exhaustive filtration 0 = M(0) ⊆M(1) ⊆M(2) ⊆ · · · such that the quotients M(i)/M(i+1)
are all free (or representable) A modules1.
The objects of Ddg(A) are by definition all semi-free right A modules. The complex NDdg(A)M is
the usual complex HomA(M,N) of A linear homomorphisms. We also use the notation RHomA(M,N)
for this complex, and the notation Ext∗A(M,N) = H∗(RHomA(M,N)) for its homology. The homotopy
category D(A) = H0(Ddg(A)) is by definition the derived category of A. It has the structure of a
triangulated category.
1The filtration can be indexed by N or by any ordinal. The two definitions are equivalent as a filtration by an ordinalcan always be transformed into one by N with the same property.
Chapter 2. Background 14
We also write Ddgperf(A) for the full dg subcategory of Ddg(A) on those dg modules which are isomor-
phic in D(A) to an object of Dperf(A). Here Dperf(A) = thickD(A)(A) is the perfect derived category
of A, that is, the smallest full, triangulated subcategory of D(A) which contains A and is closed under
summands. In general, when M is a set of objects in D(A), we denote by thinkD(A)(M) the smallest
full, triangulated subcategory of D(A) which contains the objects in M and is closed under summands.
In general, when M and N are not necessarily semi-free, RHomA(M,N) is defined to be HomA(F,N)
where F is a semi-free right A module with a chosen surjective A-linear quasi-isomorphism F'−→M . The
complex RHomA(M,N) is defined up to quasi-isomorphism, but often we will have a specific semi-free
resolution for M in mind. Moreover RHomA(M,N) is naturally a dg algebra using HomA(F, F ) as a
model with composition as product. This dg algebra depends on F only up to quasi-isomorphism of dg
algebras.
Suppose that M and N are right and left A modules respectively. The derived tensor product M⊗LAN
is similarly defined to be F ⊗AN , where F'−→M is a semi-free resolution (one could equally well resolve
N instead).
Suppose that M is a B − A bimodules and N is a C − A bimodule, for three small dg categories
A,B and C. Assume also that each xM and yN are semi-free right A modules. Then RHomA(M,N)
is naturally a C − B bimodule with yRHomA(M,N)x = RHomA(xM, yN). In particular when B = C
and M = N the quiver RHomA(M,M) is naturally a dg category with the same set of objects as B.
When M is a set of right A modules RHomA(M,M) is by definition the full dg subcategory of Ddg(A)
on M . This canonically makes M into a RHomA(M,M)−A bimodule.
We make a definition which will be useful later. Suppose that A and B are dg categories and M is
a B − A bimodule. We say that M is homologically balanced if the natural functors B → REndA(M)
and Aop → REndBop(M) are both quasi-equivalences. This condition was used by Keller in [79]. The
following is a consequence of the five lemma.
Lemma 3 ([79]). Let M be a B − A bimodule. The natural map B → RHomA(M,M) is a quasi-
equivalence if and only if the functor −⊗LB M : Dperf(B)→ D(A) is fully faithful.
2.2 Twisting Cochains and the Bar Construction
In this section we work with algebras and coalgebras over a field k. Undecorated tensor product and
Hom’s are taken over k. It is possible to work over a commutative ring instead, but we will not need
this generality.
IfA is a dg algebra with productm : A⊗A→ A and C is a dg coalgebra with coproduct ∆ : C → C⊗Cthen the complex Hom(C,A) becomes a dg algebra with the cup product f ^ g = m (f ⊗ g) ∆. This
is verified in [90, proposition 2.1.2]. We call this the convolution algebra connecting C and A. Using
this structure, an element τ ∈ Hom(C,A)−1 is called a twisting cochain if it satisfies the Maurer-Cartan
equation
∂(τ) + τ ^ τ = 0.
Assume further that both C and A are augmented. Then a twisting cochain τ : C → A is called
augmented if it vanishes on both augmentations, so the two compositions k → Cτ−→ A and C
τ−→ A→ k
are zero. The set of augmented twisting cochains from C to A is denoted tw(C,A).
Chapter 2. Background 15
Let Alg be the category of augmented dg algebras over k, and let coAlg be the category of cocomplete
augmented dg coalgebras over k. The set of twisting cochains is natural in both arguments, so we have
a functor
tw : coAlgop ×Alg→ Sets.
Theorem 2 ([90, section 2.2]). This functor is representable in both arguments: there are functors
Ω : coAlg→ Alg and B : Alg→ coAlg and natural isomorphisms
Alg(ΩC,A) ∼= tw(C,A) ∼= coAlg(C,BA).
The functors B and Ω are called the bar and cobar constructions respectively. By definition we have
an adjunction
B : Alg coAlg : Ω.
According to [90], this can be upgraded to a Quillen equivalence of model categories, this sums up much
of the force behind Koszul duality. Classical Koszul duality concerns what can be said about formality
on either side of this adjunction. This will be discussed in section 2.3.
The bar construction goes back to [42] while the cobar construction goes back to [1]. Both originate
in algebraic topology, and they remain indispensable there and in algebra. We will quickly explain how
to build the bar construction of an augmented algebra now, since this is all we need below. Many more
details, including the cobar construction, can be found in [90, section 2.2].
Let A be an augmented dg algebra. Denote A = coker(k → A), then using the augmentation there
is an induced non-unital product m : A ⊗ A → A. As a graded coalgebra BA is the tensor coalgebra
Tco(ΣA). It is traditional to write elements of BA using the notation Σa1 ⊗ · · · ⊗ Σaw = [a1| · · · |aw]. We
also write πw for the projection BA→ BwA = (ΣA)⊗w, and define two degree −1 maps b1 and b2 by the
commutativity of the following two diagrams
A A A⊗2
A
B1A B1A B2A B1A.
∂
Σ Σ
m
Σ⊗Σ Σ
−b1 b2
Finally, because BA is cofree on B1A, there is a unique coderivation b : BA → BA such that π1b =
b1π1 + b2π2 by lemma 2. Explicitly, b is given by
b[a1| · · · |aw] =
w∑i=1
(−1)εi [a1| · · · |∂(ai)| · · · |aw] +
w∑i=2
(−1)εi [a1| · · · |m(ai−1, ai)| · · · |aw]
where εi = i+ |a1|+ ...+ |ai−1|. The condition b2 = 0 is a consequence of the fact that A is an associative
dg algebra. Thus BA with b as a differential is a cocomplete dg coalgebra. This is the bar construction
of A. The universal twisting cochain into A is π : BAπ1−→ ΣA → ΣA. The bijection in the above theorem
takes a coalgebra map φ : C → BA to the twisting cochain πφ, and it follows that every other twisting
cochain into A factors uniquely through π.
The construction of ΩC is linearly dual: as an algebra ΩC = T(Σ−1C) and the differential encodes
precisely the coproduct and differential on C. When C is concentrated in degrees ≥ 2, or in degrees ≤ 0,
there is a canonical isomorphism B(C∨) ∼= (ΩC)∨. By the theorem there is also a universal twisting
Chapter 2. Background 16
cochain from C, which we denote ι : C → ΣΩC.
Quasi-isomorphisms of dg coalgebras are poorly behaved. Instead, a homomorphism φ : C → C ′
between cocomplete dg coalgebras is called a weak equivalence if the induced algebra homomorphism
Ωφ : ΩC → ΩC ′ is a quasi-isomorphism. In [85] it is proven that a weak equivalence is a quasi-
isomorphism, but this notion is stronger. Since we will not make serious use of this notion, we refer to
[90] or [85] for discussion.
Now we explain how twisting cochains can be used to build resolutions and compute derived functors.
These twisted tensor products and twisted convolution algebras will be important throughout this thesis.
If τ : C → A is a twisting cochain we may perturb the usual differential on Hom(C,A) to ∂ +
[τ,−]. The condition (∂ + [τ,−])2 = 0 is equivalent to the Maurer-Cartan equation. Equipped with
this differential Hom(C,A) is called the twisted convolution algebra associated to τ , and it is denoted
Homτ(C,A). Naturality of twisting cochains for algebra and coalgebra maps extends to naturality of
twisted convolution algebras. For later reference we need the following lemma.
Lemma 4. If τ : C → A is an augmented twisting cochain from a cocomplete dg coalgebra, and φ : A→A′ is a quasi-isomorphism dg algebras, then the induced map
φ∗ : Homτ(C,A)→ Homφτ(C,A′)
is also a quasi-isomorphism of dg algebras.
The lemma follows easily from lemma 1 using the filtration on twisted convolution algebras induced
from the the primitive filtration on C, which is complete since C is cocomplete. The twists [τ,−] and
[φτ,−] decrease the primitive filtration, so they vanish after passing to the associated graded complexes
and then then result is a standard fact about hom complexes.
Let M be a right C comodule and let N be a left A-module. The convolution algebra Hom(C,A)
acts on the left of M ⊗N via the cap product
α _ − : M ⊗N ∆⊗1−−−→ N ⊗ C ⊗M 1⊗α⊗1−−−−→ N ⊗A⊗M 1⊗m−−−→ N ⊗M.
If τ : C → A is a twisting cochain the ∂M ⊗ 1 + 1 ⊗ ∂N + (τ _ −) is a square zero differential on
M ⊗N . With this differential, it is easy to verify that the twisted tensor product M ⊗τN becomes a left
dg module over the twisted convolution algebra Homτ(C,A).
Similarly, if M is a right A module and N is a left comodule, the right action of Hom(C,A) on M⊗Nallows us to equip it with a differential ∂M ⊗1 + 1⊗∂N − (−_ τ). This twisted tensor product M ⊗τNis a right dg module over Homτ(C,A).
The following theorem is fundamental in the theory of twisting cochains.
Theorem 3 ([72, 85, 90]). If τ : C → A is an augmented twisting cochain the following are equivalent:
1. ΩC → A is a quasi-isomorphism;
2. C → BA is a weak equivalence;
3. using the counit of C, the natural A bimodule map A⊗τC ⊗τA→ A is a quasi-isomorphism.
Chapter 2. Background 17
We say that τ is acyclic if it satisfies the equivalent conditions of the theorem. Note that according
to the theorem, the universal twisting cochains Cι−→ ΩC and BA
π−→ A are both acyclic.
Note that the twisted part of the differential on A⊗τC ⊗τA decreases the primitive filtration on C.
Since C is cocomplete, this filtration witnesses the fact that A⊗τC ⊗τA→ A is a semi-free dg module
resolution over Aop ⊗A.
In the thesis of Lefevre-Hasegawa the conditions of the theorem are also shown to be equivalent to
having a canonical equivalence of derived categories − ⊗τ A : Dco(C) D(A) : − ⊗τ C, see [85] for
details.
Theorem 4. If τ : C → A is an acyclic twisting cochain then there is a canonical quasi-isomorphism
C∨'−−→ RHomA(k, k).
Proof. By the previous theorem we have a semi-free bimodule resolution A⊗τC ⊗τA '−→ A. Tensoring
down on the left produces a semi-free resolution of right dg modules C ⊗τ A '−→ k (this is sometimes
called the Koszul complex of A).
There is a natural left action C∨ on C⊗τA using the coproduct of C. This produces a homomorphism
of dg algebras φ : C∨ → HomA(C ⊗τA,C ⊗τA). Composing φ with the quasi-isomorphism HomA(C ⊗τ
A,C⊗τA)'−→ HomA(C⊗τA, k) ∼= C∨ yields the identity on C∨, and therefore φ is a quasi-isomorphism.
Example 1. Let V be a graded vector space which is degree-wise finite dimensional and concentrated
in non-negative degrees. Let S = kJV K = Sym(V ) be the free complete strictly graded commutative
algebra on V . Also let Γ = Symco(ΣV ) be the cofree strictly graded commutative coalgebra on ΣV , which
may be defined by dualising Γ ∼= (Sym((ΣV )∨))∨. This coalgebra is naturally isomorphic to the free
divided power algebra k〈ΣV 〉 on ΣV , which we will discuss in section 2.6.
The composition τ : Γ ΣV → ΣS is an acyclic twisting cochain. The Maurer-Cartan equation is
easy to check, and the twisted tensor product S ⊗τΓ⊗τ S is isomorphic to
kJV, V K〈ΣV 〉 with ∂(Σv) = v ⊗ 1− 1⊗ v for v ∈ V.
This is the classical Tate model for the diagonal which will be discussed in section 2.6. It is well known
that kJV, V K〈ΣV 〉 → kJV K is a quasi-isomorphism (see for example [44, proposition 1.9]). This example
will be generalised by theorem 13.
2.3 Koszul Duality
In this section we discuss some aspects of Koszul Duality for associative algebras. Of course, much more
can be said, and we will have to completely omit many important aspects of this theory. Continuing
from the previous section we work over a field k. All algebras are k algebras, and undecorated tensor
product and Hom’s taken over k.
In its simplest algebraic incarnation Koszul duality concerns the relation between an associative,
augmented (possibly graded or dg) algebra A, and the Yoneda algebra Ext∗A(k, k). Theorem 4 provides
the connection with twisting cochains: if τ : C → A is an acyclic twisting cochain then there is a canonical
Chapter 2. Background 18
isomorphism H∗(C∨) ∼= Ext∗A(k, k) of graded algebras. It is no secret that passing to homology (usually)
throws away information. With this in mind we say that the dg algebra RHomA(k, k) is Koszul dual to
A. We use the notation A! = RHomA(k, k) with the understanding that this is only defined up to quasi-
isomorphism. Since π : BA→ A is an acyclic twisting cochain (BA)∨ functorially provides a model for
the Koszul dual dg algebra to A. Note that this model is augmented: we should really consider A! to
be an augmented dg algebra defined up to augmented quasi-isomorphism.
Koszul duality has many algebraic and geometric incarnations, and is best motivated through exam-
ples. For instance Priddy’s original work [104] discusses in detail the Steenrod algebra from algebraic
topology. But there are two basic reasons why one should be interested in A! if one is interested in A.
Firstly, the representation theory of A is closely related to that of A!. This goes back at least to [26],
and will be explained in part below. Secondly, under mild finiteness conditions it is possible to recover
A from A!. In fact there is a canonical quasi-isomorphism A!! ' A. We will sketch the reason for this
now, much more detail can be found in [90] or [85].
The price of working with algebras is that repeated dualisation necessitates some finiteness conditions.
The theory of twisting cochains suggests a version of Koszul duality using coalgebras which avoids
dualising unnecessarily. Details of this picture are worked out in the thesis of Lefevre-Hasegawa [85].
For readability we mostly stick to the algebra side in this thesis. Thus, let us say that an augmented dg
algebra A is strongly connected if H∗(A) is degree-wise finite dimensional and it is either homologically
connected or cohomologically simply-connected, which means H∗(A) must be concentrated in degrees
≥ 1 or in degrees ≤ −2 respectively.
Theorem 5. If A is a strongly connected then A!! is canonically quasi-isomorphic to A. Thus, A can
be recovered from A! up to quasi-isomorphism.
A inductive argument shows that up to quasi-isomorphism we can arrange that A itself is degree-
wise finite dimensional and concentrated in degrees ≥ 1 or in degrees ≤ −2 (a more intuitive way of
doing this is to replace A with a minimal A∞ model using theorem 8, this will be explained in the next
section). Having done this, there is a canonical isomorphism (B(BA)∨)∨ ∼= (ΩBA)∨∨ ∼= ΩBA. One can
show that the natural homomorphism ΩBA → A is always a quasi-isomorphism by writing down an
explicit contraction as in [85, lemme 1.3.2.3] (note that this is also a consequence of theorem 3). Since
(B(BA)∨)∨ is a model for the double Koszul dual to A, this produces a quasi-isomorphism A!! ' A.
Since A! → RHomA(k, k) is a quasi-isomorphism by definition, the theorem may be restated in the
terminology of section 2.1.
Corollary 1. If A is a strongly connected then k is homologically balanced as an A! −A bimodule.
The following is perhaps the most basic derived equivalence which Koszul duality yields. These ideas
go back to [26]. With appropriate definitions the theorem extends to the A∞ setting, worked out in [92].
Lefevre-Hasegawa establishes a much larger and more generally applicable derived equivalence in [85],
relating A and its Koszul dual dg coalgebra (the source of any acyclic cochain into A).
Recall that Dperf(A) = thickD(A)(A) is the prefect derived category of A. We also write Dfd(A) for
the full subcategory of D(A) on those objects with finite dimensional homology.
Theorem 6. If A is strongly connected we have natural equivalences of triangulated categories
Dfd(A) ' Dperf(A!) and Dperf(A) ' Dfd(A!).
Chapter 2. Background 19
Proof. The first equivalence is quite formal using the techniques of Keller [75], and strong connectedness
is not needed.
First note that Dfd(A) = thickD(A)(k). Let K'−→ k be a semi-free right A module resolution.
Following [75] we work with the dg enhancement Ddgfd (A) of semi-free dg modules with finite dimensional
cohomology. Since k is generator for Dfd(A) the functor
RHomA(k,−) = HomA(K,−) : Ddgfd (A)→ Ddg
perf(EndA(K))
is a quasi-equivalence by the five lemma. Then since A! ' EndA(K) we also have an equivalence
Dperf(EndA(K)) ' Dperf(A!).
The second equivalence then follows by reversing the roles of A and A! using theorem 5.
Classical Quadratic Koszul Duality
Starting with an augmented graded algebra A, it would be desirable to be able to replace the large dg
algebras RHomA(k, k) or (BA)∨ with the simpler graded algebra Ext∗A(k, k) in theorems 5 and 6. With
this in mind, we define A to be Koszul if A is strongly connected and A! is formal, which mean there
is a chain of quasi-isomorphisms connecting A! to H∗(A!) = Ext∗A(k, k). We will explain here how this
relates to the more classical picture of Koszul duality, which concerns algebras and coalgebras defined
by quadratic data, and small resolutions over them. These ideas go back to the seminal work of Priddy
[104].
Miraculously, the assumption that A is Koszul leads (with the data of a presentation of A) to a
simple, canonical quasi-isomorphism (BA)∨'−→ Ext∗A(k, k), or more fundamentally TorA∗ (k, k)
'−→ BA.
So, let A be an augmented graded algebra and suppose we have a graded vector space V with a map
V → A such that T(V ) → A is surjective. Assume that V is minimal, meaning that the kernel of this
map lies in T≥2(V ). Choose also a minimal space of relations R ⊆ T≥2(V ), so that A is the quotient of
T(V ) by the two-sided ideal (R) generated by R. Minimality means here that (V ·R+R ·V )∩R = 0. 2
By choosing a space of generators any semi-free dg module can be written in the form M = (U⊗A, ∂)
where U =⋃U(i) is a filtered vector space such that U(0) = 0 and ∂(Ui) ⊆ Ui−1 ⊗ A. We say that M
(equipped with this space of generators) is linear if ∂(U) ⊆ U ⊗ V . The minimal semi-free resolution of
the augmentation module k begins
· · · → Σ2R⊗A→ ΣV ⊗A→ A→ 0.
And therefore TorA∗ (k, k) will begin (k + ΣV + Σ2R + ...). Note that for this resolution to be linear R
must be quadratic.
Aside from the algebra A = A(V,R) = T(V )/(R), there is an equally natural way to define a coalgebra
from quadratic data of V and R ⊆ V ⊗2. We can take the cocomplete, augmented coalgebra C(V,R)
which is cofree on V subject to the constraint that C(V,R) → C(V,R) ⊗ C(V,R) → V ⊗ V factors
2With this data we can build the dg algebra T(V, ΣR) with ∂(Σr) = r ∈ T(V ). This is the first step in the non-commutative Tate resolution of A = T(V )/(R): as in section 2.6 one inductively constructs a dg algebra resolution
T(V0, V1, V2, ...)'−→ A with V0 = V and V1 = ΣR and Vi in degree i. We can tautologically interpret T(V0, V1, ...) = ΩC as
the cobar construction of an A∞ coalgebra, which comes with an acyclic twisting cochain C → A (A∞ structures will bediscussed shortly). It follows from this discussion that there is a canonical isomorphism C = (k+ΣV0+ΣV1+...) ∼= TorA∗ (k, k).If A is Koszul, we can construct C with no higher coproducts. In other words the differential T(V0, V1, ...) must be quadratic,and in particular we must be able to choose the relations R quadratic.
Chapter 2. Background 20
through R → V ⊗ V . Concretely, C(V,R) may be presented as the sub-coalgebra of Tco(V ) whose
weight w part is Cw =⋂p+2+q=w V
⊗p ⊗ R ⊗ V ⊗q. It is easy to check that there is a twisting cochain
τ : C(ΣV, Σ2R) ΣV → ΣA. This factors through the universal one via the inclusion C(ΣV, Σ2R) →Tco(ΣV ) → BA.
Theorem 7. Let A = T(V )/(R) be as above, and assume A is strongly connected. The following are
equivalent:
1. the dg algebra A! = RHomA(k, k) is formal;
2. the augmentation module k admits a linear resolution;
3. R can be chosen quadratic and C(ΣV, Σ2R) → BA is a quasi-isomorphism, so that C(ΣV, Σ2R) ∼=TorA∗ (k, k).
4. R can be chosen quadratic and the twisting cochain τ : C(ΣV, Σ2R)→ A is acyclic.
An algebra satisfying these equivalent conditions is called Koszul.
The equivalence of 3 with 2 is in [26]. Under the strongly connected assumption 4 and 3 are easily
seen to be equivalent. The obtained resolution C(ΣV, Σ2R) ⊗τA ∼−→ k is called the Koszul complex by
Beilinson, Ginzburg and Soergel. It generalises the classical Koszul complex, when A is a polynomial
algebra (see example 1). These results ultimately trace back to Priddy [104]. The connection to 1 is in
Keller’s [77], but see also earlier the work of Gugenheim and May [62].
2.4 A∞ Algebras
This section contains a quick introduction to the theory of A∞ algebras. The survey [76] is a good place
to start for more intuition, and much more detail can be found in [90]. We continue with the notation
of the previous section, working implicitly over a field k.
A∞ algebras (and A∞ spaces) were invented by Stasheff [113] in the early 60s as a tool to understand
loop spaces and their relation to general H-spaces. In short, he used them to develop a fundamental
recognition principle for loop spaces. They have since found many important applications in connection
with algebra, geometry and mathematical physics (see [78] for some references in this direction). We will
attempt to motivative this structure at a basic algebraic level as a continuation of the previous sections.
Outside of the Koszul case working with A! can be complicated. Most models, such as (BA)∨, are
extremely large and contain more information than necessary. But since A! is not formal the graded
algebra Ext∗A(k, k) does not contain enough information. Koszul algebras are surprisingly plentiful, but
non-Koszul algebras certainly arise - the quadratic condition being the main restriction.
It turns out that Ext∗A(k, k) is naturally anA∞ algebra: it has a series of operationsmn : Ext∗A(k, k)⊗n →Ext∗A(k, k) (unique up to isomorphism) which together satisfy certain higher associativity rules. This
structure solves the information problem: it is precisely enough to functorially recover A. Thus we think
of it as an A∞ model for A!. Having done this, we also solve the problem of unwieldy models. The issue
that remains is actually computing these higher products.
More motivation in a similar vein comes from looking for small resolutions. As we have have just
seen, if A is a Koszul there is an acyclic twisting cochain τ : TorA∗ (k, k)→ A. Consequently we obtain a
(potentially) small, functorial resolution M ⊗τTorA∗ (k, k)⊗τA of any right A module M . In fact this is
Chapter 2. Background 21
always possible. In general TorA∗ (k, k) will be an A∞ coalgebra, and knowing this structure one can form
the twisted tensor product M ⊗τ TorA∗ (k, k) ⊗τ A resolving any right A module. Since we won’t need
these resolutions in this thesis, we refer the reader to [90] for information on twisted tensor products in
the A∞ setting.
An A∞ algebra is a complex A with a sequence of degree n− 2 operations mn : A⊗n → A for n ≥ 2.
These should satisfy the quadratic Stasheff identities
∂(mn) +∑
r+s+t=n
(−1)rs+tmr+1+t(1⊗r ⊗ms ⊗ 1⊗t) = 0,
as well as the strict unitality conditions mn(−, ..., 1, ...,−) = 0 for n ≥ 3 and m2(−, 1) = m2(1,−) = idA.
We say that A is augmented if it is equipped with a splitting A = A+k and all of the operations preserve
A. An A∞ algebra is called minimal if its differential vanishes.
These operations together describe a structure which is associative up to homotopy: the higher
associator∑
(−1)rs+tmr+1+t(1⊗r ⊗ms⊗ 1⊗t) is a boundary in Hom(A⊗n, A). In particular, the second
Stasheff identity says ∂(m2) = 0 and the third says that m2(m2⊗ 1− 1⊗m2) = ∂(m3), therefore H∗(A)
is canonically an associative algebra. The sequence of operations can be thought of as a coherent reason
for H∗(A) to be associative. Even when A is minimal and no reason seems necessary, the operations can
witness associativity in a nontrivial way, carrying important information.
Note that differential graded algebras are exactly A∞-algebras for which mn = 0 when n ≥ 3.
An augmented A∞ algebra possesses exactly enough structure to build a bar construction. The
underlying graded coalgebra of BA is Tco(ΣA), with BwA = (ΣA)⊗w, just as in section 2.2. And as in
that section b1 = −Σ∂Σ−1 is the differential naturally induced on B1A, while bn : BnA→ B1A is defined
by the commutativity of the following diagram
A⊗n
A
BnA B1A.
mn
Σ⊗n Σ
(−1)nbn
By lemma 2 there is a unique coderivation b : BA→ BA such that π1b = b1π1 + b2π2 + b3π3 + .... The
condition b2 = 0 is equivalent to the Stasheff identities. Thus the bar construction BA is dg coalgebra.
We think of it as being Koszul dual to A. In particular the linear dual is a dg algebra and we may write
A! = (BA)∨. Through this construction augmented A∞ algebras are essentially the same thing as cofree
cocomplete augmented dg coalgebras3.
An A∞ coalgebra C may be defined dually in terms of higher coproducts ∆n : C → C⊗n satisfying
dual Stasheff relations. We skip this discussion and refer to [90] for details.
Alternatively, an augmented A∞ coalgebra structure on C = C+k can be specified by an augmented
differential d on the tensor algebra T(Σ−1C). Tautologically, T(Σ−1C) equipped with this differential is
the cobar construction ΩC. Again, minimality means that the differential of C vanishes. Equivalently,
d strictly increases weight. A dg coalgebra is then an A∞ coalgebra such that the differential of ΩC
increases weight by no more than one.
The important assumption that C is cocomplete translates to the assumption that ΩC should be
semi-free: C must admit an exhaustive filtration 0 = C(0) ⊆ C(1) ⊆ ... ⊆ C such that d(Σ−1C(i)) ⊆3It is not possible to simply drop the word “augmented” here. Removing this assumption from one side introduces
curvature on the other. This was discovered by Positselski [103].
Chapter 2. Background 22
T(Σ−1C(i−1)) for all i. For dg coalgebras this notion of cocompleteness is equivalent to the usual one.
A∞ coalgebras are often produced using a noncommutative version of the Tate construction (the
commutative version will be discussed in detail in section 2.6). In short, given a graded algebra A
of the form T(V )/(R) with R ⊆ T≥2(V ), it is possible to inductively construct a semi-free resolution
T(W )'−→ A with some differential d which presents T(W ) as the cobar construction of a cocomplete A∞
coalgebra C = k + ΣW . Some details on this construction can be found in [25].
Having a general enough theory of twisting cochains (as developed in [90]) allows one to assert that
the composition C → ΩC → A is an acyclic twisting cochain. The dual C∨ is an A∞ algebra and we
may write A! = C∨. To explain why it is acceptable to write this last equality we at least need to know
what is a morphism of A∞ algebras.
An A∞ morphism φ : A→ A′ consists of a sequence of maps φn : A⊗n → A′ for n ≥ 1 having degree
1− n respectively. They must satisfy the quadratic identities∑r+s+t=n
(−1)rs+tφr+1+t(1⊗r ⊗ms ⊗ 1⊗t) −
∑i1+...+ik=n
(−1)umk(φi1 ⊗ ...⊗ φik) = ∂(φn)
where u = (k − 1)(i1 − 1) + (k − 2)(i2 − 1) + · · · + 2(ik−2 − 1) + (ik−1 − 1). We also impose the strict
unitality conditions φn(−, ...,−, 1,−, ...,−) = 0 for n ≥ 2 and φ1(1) = 1. When A and A′ are augmented,
we say that φ is augmented if it takes each A⊗n
into A′.
The first identity asserts that φ1 is a chain map. The second asserts that φ1m2−m2(φ1⊗φ1) = ∂(φ2).
Therefore φ induces a homomorphism H∗(φ) : H∗(A) → H∗(A′) of associative algebras. The higher
components should be thought of as a coherent reason for this to be the case. We say that φ is strict if
φn = 0 for n ≥ 2.
We compose A∞ morphisms φ : A→ A′ and ψ : A′ → A′′ according to the rule
(ψφ)n =∑
i1+...+ik=n
(−1)uψk(φi1 ⊗ ...⊗ φik)
with the same sign u as above. It is not difficult to check that φ is an isomorphism (that is, it admits a
two-sided inverse) if and only if φ1 is an isomorphism of complexes. We say that φ is a quasi-isomorphism
if H∗(φ) is an isomorphism.
Let φ be an augmented morphism of augmented A∞ algebras. By cofreeness the shifted maps
Σφn(Σ−1)⊗n : BnA → B1A′ assemble uniquely into a coalgebra morphism B(φ) : BA → BA′. Being
careful of signs, one can check that the identities defining an A∞ morphism are equivalent to the assertion
that B(φ) is a chain map. Thus through the bar construction the category of augmented A∞ algebras
is equivalent to the category of cofree cocomplete dg coalgebras. Strict A∞ morphisms correspond to
coalgebra morphisms which preserve weight.
The bar-cobar formalism easily allows one to show that every A∞ algebra is quasi-isomorphic to a dg
algebra, and that any quasi-isomorphism of A∞ algebras can be transferred to strict quasi-isomorphism
between the corresponding dg algebras (there is a canonical quasi-isomorphism A→ ΩBA of A∞ alge-
bras). Thus A∞ algebras are homotopically no more general than dg algebras. They simply provide a
more flexible presentation for the same homotopical information carried by dg algebras. This perspective
is especially compelling with the following important theorem of Kadeishvilli in mind.
Theorem 8 (Kadeishivili’s Minimal Model Theorem [73]). Let A be an augmented A∞ algebra. The
associative algebra H∗(A) admits the structure of a minimal augmented A∞ algebra extending its natural
Chapter 2. Background 23
product, in such a way that there is a quasi-isomorphism
H∗(A)'−→ A
which induces the identity map in homology. Moreover, this structure is unique up to isomorphism of
A∞ algebras.
One caveat is that an isomorphism of A∞ algebras can be a relatively complicated object.
The theorem is extremely important even when applied to a dg algebra. Starting with a graded
algebra A, it produces a minimal A∞ algebra quasi-isomorphic to the Koszul dual (BA)∨. Thus there
is a natural A∞ structure on Ext∗A(k, k), and we call this the minimal model for A!. When A is strongly
connected (using the same definition as in section 2.3) a similar argument to the proof of theorem 5
produces a canonical quasi-isomorphism (BExt∗A(k, k))∨'−→ A. Thus from the A∞ algebra Ext∗A(k, k)
we can functorially recover A.
If we build a minimal non-commutative Tate model ΩC'−→ A as sketched above, then uniqueness of
minimal semi-free dg algebras allows us the make the identification ΩC ∼= (BExt∗A(k, k))∨ and conclude
that there is an isomorphism C∨ ∼= Ext∗A(k, k) of A∞ algebras.
We end this section with a very general example showing how to compute a part of the natural A∞
structure on Ext∗A(k, k). If A = T(V )/(R) with R ⊆ T≥2(V ) as above, then examining the first two
steps in the minimal non-commutative Tate construction readily yields the following theorem of Keller.
Theorem 9 (Keller [77, proposition 2]). There are canonical isomorphisms Ext1A(k, k) = Σ
−1V ∨ and
Ext2A(k, k) = Σ
−2R∨, and for every n the following diagram commutes
Ext1A(k, k)⊗n Ext2
A(k, k)
(V ∨)⊗n R∨n .
mn
Σ⊗n Σ
(−1)nι∨n
In the statement Rn is the weight n part of R and ιn : Rn → V ⊗n is the natural inclusion.
For algebras of global dimension 2 this completely describes the Koszul dual A∞ algebra Ext∗A(k, k).
2.5 Models for the Hochschild Cochain Complex
In this section we will discuss models for the Hochschild cochain complex in various settings, as well the
structure that these models enjoy.
To begin with let A be an associative algebra (possibly graded or dg) over a commutative ring k. We
denote the enveloping algebra by Aev = Aop ⊗k A. Bimodules over A are equivalent to left or right Aev
modules, and in particular A is naturally both a left and right Aev module.
Hochschild homology and cohomology were originally defined in terms of the bar construction. When
A is flat as a k module we may instead define Hochschild homology as
HH∗(A/k,A) = TorAev
∗ (A,A).
Chapter 2. Background 24
When A is projective over k we may define Hochschild cohomology as
HH∗(A/k,A) = Ext∗Aev(A,A).
We will use these definitions briefly in section 2.6 (example 3 and remark 2). However, in the rest of this
section we will discuss much more structural models for Hochschild cohomology which will be important
in chapter 4.
We continue with the conventions of the last few sections, so k is once again a field and everything
is assumed to be k linear. From this point on we elide k from our notation for Hochschild cohomology.
Let τ : C → A be a twisting cochain from an augmented cocomplete dg coalgebra to an augmented
dg algebra. Recall that twisted tensor products and the twisted convolution algebra Homτ(C,A) were
defined in section 2.2. In that section we also observed that A ⊗ C ⊗ A is a bimodule over Hom(C,A)
using the so-called cap product. The twisted tensor product A⊗τC⊗τA by definition has the differential
∂A ⊗ 1⊗ 1 + 1⊗ ∂C ⊗ 1 + 1⊗ 1⊗ ∂A + (τ _ −)− (−_ τ).
It is easy to check that this makes A ⊗τ C ⊗τA a dg bimodule over over Homτ(C,A). It is also clear
that this action commutes with the A bimodule structure. In particular, the left action gives us a
homomorphism Homτ(C,A)→ EndAev(A⊗τC ⊗τA) of dg algebras.
At this point we can make an argument similar to the proof of theorem 4. There is a natural bimodule
homomorphism A⊗τC ⊗τA→ A which gives rise to the second map below
Homτ(C,A)→ EndAev(A⊗τC ⊗τA)→ HomAev(A⊗τC ⊗τA,A).
Note that the composition from left to right is an isomorphism of complexes (this is clear on the level
of vector spaces and the differentials are automatically compatible, no calculation necessary).
Finally, if we assume τ is acyclic then according to theorem 3 A⊗τC⊗τA→ A is a quasi-isomorphism.
Then EndAev(A⊗τC ⊗τA)→ HomAev(A⊗τC ⊗τA,A) is a quasi-isomorphism as well. Cocompleteness
of C is essential here because this guarantees that A⊗τC ⊗τA is semi-free as a dg bimodule. Thus, we
have constructed a canonical dg algebra quasi-isomorphism
Homτ(C,A)'−→ EndAev(A⊗τC ⊗τA) = RHomAev(A,A).
In particular, it is a consequence of theorem 3 that the Hochschild cohomology HH∗(A,A) can be com-
puted as H∗(Homτ(C,A)) using any acyclic twisting cochain into A4. This description of Hochschild
cohomology is essentially in [90] or [85]. Negron has used this description to compute Hochschild co-
homology in [102] (the author also deals more generally with non-augmented algebras and curved dg
coalgebras). Similar remarks apply to Hochschild cohomology with coefficients in a bimodule; details on
this can also be found in op. cit. This description of Hochschild cohomology has been used as well in
topology by Berglund and Borjeson [29].
Example 2. An obvious application is the computation of Hochschild cohomology for Koszul algebras.
4Note that if we had used the right cap action of Homτ(C,A) on A ⊗τ C ⊗τ A then the same argument would haveproduced an anti-isomorphism H∗(Homτ(C,A)) ∼= HH∗(A,A)op. Both induce the same map on cohomology since they areequalised by EndAev (A⊗τC ⊗τA)→ HomAev (A⊗τC ⊗τA,A). So, we have a funny proof that Hochschild cohomology iscommutative.
Chapter 2. Background 25
If A = T(V )/(R) is a Koszul algebra as in section 2.3 and C = C(ΣV, Σ2R) then we obtain a small model
Homτ(C,R) for the Hochschild cochain complex of A using theorem 7.
For example, let S = Sym(V ) (assumed strongly connected to fit into the framework of section 2.3,
otherwise we should use the completion of S) and Γ = Symco(ΣV ) be as in example 1. Then τ : Γ→ ΣV →ΣS is an acyclic twisting cochain. Since S and Γ are commutative and cocommutative (this is extremely
rare, it can essentially only occur in this example) the twist on Homτ(Γ, S) = Hom(Γ, S) vanishes. We
obtain a graded version of the Hochschild-Kostant-Rosenburg theorem HH∗(S, S) ∼= Hom(Γ, S). The
original version of the Hochschild-Kostant-Rosenburg theorem for smooth algebras will appear in remark
2 below.
Let A be an augmented dg algebra. The universal twisting cochain π : BA→ A is acyclic according
to theorem 3, and we set
C∗(A,A) = Homπ(BA,A).
This is exactly the classical reduced Hochschild cochain complex: it has a weight decomposition Cw(A,A) ∼=Hom(A
⊗w, A) and unraveling the definitions recovers Hochschild’s original differential. The Hochschild
cohomology of A is by definition HH∗(A,A) = H∗(C∗(A,A)). The convolution product is exactly the
classical cup product. We will discuss shortly how to see Gerstenhaber’s Lie bracket in terms of the bar
construction.
The Hochschild cochain complex has a natural decreasing filtration which we call called the weight
filtration
C(w)(A,A) = Homπ(B≥wA,A).
Note that this filtration is complete since BA is cocomplete. The induced weight filtration on Hochschild
cohomology is denoted HH(n)(A,A).
Hochschild Cohomology for A∞ Algebras
Hochschild cohomology for A∞ algebras can be defined using an A∞ analogue of the twisted convolution
algebra. This construction is essentially contained in [85, chapitre 8], but we follow the approach of [34,
section 2.3], where more detail can be found.
Let C be an augmented cocomplete coalgebra and A an augmented A∞ algebra. The complex
Hom(C,A) inherits a sequence of convolution operations
Mn(f1, ..., fn) = mn(f1 ⊗ ...⊗ fn)∆(n).
One verifies as in [85, lemma 8.1.1.4] that these operations make Hom(C,A) an A∞ algebra. Moreover,
the augmentations of C and A canonically make Hom(C,A) an augmented A∞ algebra.
A twisting cochain C → A is a degree −1 map τ : C → A which satisfies the higher Maurer-Cartan
equation
∂(τ) +∑n≥2
Mn(τ, ..., τ) = 0
The sum makes sense since C is cocomplete. We say that τ is augmented if it vanishes on both augmen-
tations, exactly as in section 2.2.
By cofreeness of the bar construction there is an augmented coalgebra homomorphism φτ : C → BA
lifting τ . By construction the higher Maurer-Cartan equation holds if and only if φτ respects the
Chapter 2. Background 26
differentials of these coalgebras. It follows that there is a natural bijection
tw(C,A) ∼= coAlg(C,BA)
where tw(C,A) is the set of augmented twisting cochains from C to A. In particular the projection
π : BA→ A is an augmented twisting cochain, and all other augmented twisting cochains into A factor
uniquely through π.
From an augmented twisting cochain τ : C → A one can form the twisted convolution A∞ algebra
Homτ(C,A). The reader should be warned that all of the operations on Hom(C,A) should be twisted to
build this structure, not only the differential. This is done in [85, chapitre 8] using twists of topological
A∞ algebras. A simpler construction is given in [34, section 2.3].
For simplicity we present here only the formula which shows how to twist the differential. We use
the higher commutators [−;−]p,q : A⊗p ⊗A⊗q → A. These are the degree 2− (p+ q) maps given by
[v1, ..., vp; vp+1, ..., vp+q]p,q =∑
σ∈sh(p,q)
(−1)|σ|(−1)|σ;v|mp+q(vσ−1(1) ⊗ ...⊗ vσ−1(p+q))
where sh(p, q) is the set of permutations in Sp+q which separately preserve the order of 1, ..., p and
of p + 1, ..., p + q. Note that [−;−]1,1 is the usual commutator for m2. Note also that [−;−]p,qt =
(−1)pq[−;−]q,p, where t is the natural isomorphism A⊗p ⊗ A⊗q → A⊗q ⊗ A⊗p which acts according to
the Koszul sign rule.
The twisted differential on Hom(C,A) is defined using higher commutators for the convolution op-
erations. It is given by
∂τ = ∂ +∑n≥1
[τ, ..., τ ;−]n,1.
Unraveling this to an explicit formula yields
∂(f) +∑i,j
(−1)(n+1)jmi+1+j(τ⊗i ⊗ f ⊗ τ⊗j)∆(i+1+j)
for the boundary of an element f in Homτ(C,A)n. The verification that (∂τ )2 = 0 is contained in [34,
section 2.3]. It is easiest to first connect this definition the the coderivation model introduced below.
Finally, the Hochschild cochain complex of A is C∗(A,A) = Homπ(BA,A). Its homology is the
Hochschild cohomology HH∗(A,A). In this context we also have a complete weight filtration with the
same definition
C(w)(A,A) = Homπ(B≥wA,A).
The induced weight filtration on Hochschild cohomology is denoted HH(n)(A,A) as before. A construc-
tion of the natural A∞ algebra structure on C∗(A,A) can be found in [34, section 2.3].
The Gerstenhaber Lie algebra
In his fundamental investigations into the deformation theory of associative algebras [52], Gerstenhaber
introduced a degree 1 Lie bracket on HH∗(A,A). This structure together with the cup product (and
their compatibility) is now summed up by saying that HH∗(A,A) is a Gerstenhaber algebra.
In this section we discuss another description of the twisted hom complex which makes the definition
Chapter 2. Background 27
of Gerstenhaber’s bracket transparent. This description goes back to Getzler and Jones [54] and Stasheff
[114].
A morphism ξ : C ′ → C of dg coalgebras naturally makes C ′ into a bicomodule over C, so we may
consider the space coDer(C ′, C) of coderivations from C ′ into C, which was introduced in section 1.2.
This is the space of linear maps C ′ → C which satisfy the (dual) Leibniz identity ∆q = (ξ⊗ q+ q⊗ ξ)∆.
It has a natural differential as a sub-complex of Hom(C ′, C).
Now let τ : C → A be an augmented twisting cochain from an augmented cocomplete dg coalgebra
to an augmented A∞ algebra, and let φ = φτ be its lift to a coalgebra morphism C → BA.
By lemma 2, there is a canonical isomorphism of graded vector spaces
Hom(C,A) ∼= Σ−1Hom(C, ΣA)⊕ C∨ ∼= Σ
−1coDer(C,BA)⊕ C∨.
At the same time, there is a natural chain map ad : C∨ → coDer(C,BA) which sends f : C → k to the
inner coderivation
ad(f) : C∆−−→ C ⊗ C φ⊗f−f⊗φ−−−−−−−→ BA.
The cone of this map is denoted
cone(ad) = Σ−1coDer(C,BA) o C∨.
As promised above, the verification that (∂τ )2 = 0 is achieved as follows.
Proposition 1. The isomorphism Homτ(C,A) ∼= Σ−1coDer(C,BA) o C∗ is one of complexes, and in
particular (∂τ )2 = 0.
This is essentially due to Getzler and Jones, and in the dg situation Stasheff [114]. The computation
can be found for instance in [34, proposition 2.23].
In the case of the identity homomorphism C → C the complex coDer(C,C) is naturally a dg Lie
algebra. The bracket is the graded commutator [p, q] = pq − (−1)|p||q|qp, making coDer(C,C) a sub Lie
algebra of the associated Lie algebra End(C)Lie.5
The natural action of coDer(C,C) on C makes coDer(C,C) o ΣC∨ a dg Lie algebra as follows. The
bracket on coDer(C,C) has just been defined, while if f, g ∈ C∨ and q ∈ coDer(C,C) then
[Σf, q] = −(−1)|q|(|f |+1)[q, Σf ] = Σ(fq) and [Σf, Σg] = 0.
In particular, if A is an augmented A∞ algebra then
ΣC∗(A,A) = ΣHomπ(BA,A) ∼= coDer(BA,BA) o Σ(BA)∨
is naturally a dg Lie algebra.
Proposition 2. This bracket on ΣC∗(A,A) canonically gives HH∗(A,A) the structure of a Gerstenhaber
algebra. It coincides with the classical one introduced by Gerstenhaber.
5This is a graded Lie algebra (with compatible differential) in the full sense of section 2.8. The reduced square is givenby q[2] = q2 on odd elements.
Chapter 2. Background 28
The appropriate chain level statement is that C∗(A,A) is a B∞ algebra (also known a brace algebra,
or E2 algebra, or G∞ algebra), this is a version of the so-called Deligne conjecture (now proven by a
number of authors [115, 96, 119]). For information on the B∞ structure of C∗(A,A) see [79]. The
first statement of the proposition is a consequence of this fact. The second statement is a calculation
contained in [54].
Finally, we remark that there is also a dual picture for the Hochschild cohain complex in terms of
derivations out of the cobar construction. We refer to [34, section 2.3] for details.
Naturality of the Hocschild Cochain Complex
Naturality for twisting cochains extends readily to the A∞ situation. Let τ : C → A be an augmented
twisting cochain. Any dg coalgebra morphism ξ : C ′ → C yields a twisting cochain τξ : C ′ → A in the
obvious way, while a morphism ψ : A → A′ of A∞ algebras yields a twisting cochain ψτ : C → A′ by
composing
Cφτ−−→ BA
B(ψ)−−−→ BA′π−→ A′.
Now suppose that φ : A → A′ and ψ : A′ → A′′ is a sequence of A∞ morphisms. Denote the corre-
sponding twisting cochains by τ = πB(φ) and σ = πB(ψ), and note that ψτ = σB(φ). We obtain two
maps
Homτ(BA,A′)ψ∗−−−→ Homψτ(BA,A′′) and Homσ(BA′, A′′)
φ∗−−−→ HomσB(φ)(BA,A′′)
defined as follows. Firstly φ∗ = Hom(φ,A′) is the obvious map. We define ψ∗ in terms of coderivations:
Σ−1coDer(BA,BA′) o (BA)∨
ψ∗−−−→ Σ−1coDer(BA,BA′′) o (BA)∨
is given by ψ∗ = Σ−1coDer(BA,B(ψ)) o (BA)∨. We state a generalisation of lemma 4.
Proposition 3. Both φ∗ and ψ∗ are chain maps. Furthermore, φ∗ and ψ∗ are quasi-isomorphisms
whenever φ and ψ are respectively.
The proof comes down to another application of lemma 1, the key point being that the twisted part
of the differential on Homτ (BA,A′) decreases the weight filtration, and that gr(φ) : gr(BA)→ gr(BA′)
is a quasi-isomorphism whenever φ is by the Kunneth theorem. Details can be found in [34, proposition
2.27].
Suppose that φ : A→ A′ is a quasi-isomorphism of A∞-algebras corresponding to a twisting cochain
τ = πB(φ). An immediate consequence of the proposition is that we have a canonical chain of quasi-
isomorphisms
C∗(A′, A′)φ∗−−−→ Homτ(A,A′)
φ∗←−−− C∗(A,A).
And in particular a canonical isomorphism HH∗(A′, A′) ∼= HH∗(A,A).
If φ and ψ are strict morphisms of dg algebras then φ∗ and ψ∗ are easily seen to be morphisms of dg
algebras. In general, the twisted convolution algebras inherit A∞ structures and φ∗ is a strict morphism
of A∞ algebras, while ψ∗ extends canonically to a possibly non-strict morphism of A∞ algebras (see [34,
section 2.3]).
Chapter 2. Background 29
The Projection and Shearing Morphisms
On the Hochschild cohomology of any dg or A∞-algebra A we have two well known maps
Π : HH∗(A,A)→ H(A) and χ : HH∗(A,A)→ H(A!) = ExtA(k, k)
which we call the projection and shearing morphisms respectively. On the chain level the definitions
are very simple: Π : Homπ(BA,A)→ Hom(k,A) = A is defined by composing with the coaugmentation
k → BA while χ : Homπ(BA,A) → Hom(BA, k) = A! is defined by composing with the augmentation
A→ k.
For ordinary graded algebras, the image of the projection map HH∗(A,A)→ A is exactly the graded
centre Z(A). Section 4.2 will be devoted to discussing the meaning of the image in general. Shamir [110]
discusses the shearing map in detail.
Note that the projection morphism is the quotient by the positive part of the weight filtration. In
[34, sections 2.3 and 3.4] it is explained that Hochschild cohomogy naturally has another filtration which
is in some sense Koszul dual to the weight filtration. We call this the shearing filtration because the
shearing morphism is the quotient by the positive part of the shearing filtration.
As a special case of the naturality from proposition 3, we get naturality of the projection and shearing
maps for quasi-isomorphisms of A∞-algebras. That is, if φ : A → A′ is a quasi-isomorphism of A∞-
algebras, then we have a diagram
(BA)∨ C∗(A,A) A
' •
(BA′)∨ C∗(A′, A′) A′
Πχ
' φB(φ)∨ '
Πχ
(2.1)
which commutes after taking homology.
In terms of coderivations the shearing map is the obvious quotient
C∗(A,A) = s−1coDer(BA,BA) o (BA)∨ → (BA)∨,
and there is a similar interpretation of the projection map in terms of derivations.
Hochschild Cohomology for (Non-Augmented) Categories and Functoriality
Hochschild cohomology is also defined for small dg andA∞ categories (this is sometimes called Hochschild-
Mitchell cohomology after [100]). If A is augmented, the definition C∗(A,A) = Homπ(BA,A) continues
to make perfect sense. In this situation the bar construction BA is cofree as a cocomplete cocategory
with the same objects as A. The composition gives rise to a differential on BA and the contents of this
section so far above generalises verbatim to this situation. Rather than explain this properly we refer
the reader to [80].
However, when working with categories the non-augmented situation will often arise by force, since
module categories are never augmented. So, we collect here some definitions for later use, restricting to
the dg case for simplicity.
Chapter 2. Background 30
Fix a small dg category A, and consider its unreduced bar construction
BunrA = Tco(ΣA) =⊕w≥0
(ΣA)⊗w =⊕w≥0
x0,...,xw
Σ(xwAxw−1)⊗ · · · ⊗ Σ(x1
Ax0),
the cofree graded cocategory on the graded quiver ΣA. The obvious “bar” dg structure on BunrA is
problematic (it is contractible), but we can still make sense of the differential on the A-bimodule A⊗π
BunrA⊗πA. We will be content cheating and simply defining it by the following formula: the boundary
∂π(f) of an element f = g ⊗ [f1|...|fw]⊗ h is the sum
∂(f) + (−1)|g|+1gf1 ⊗ [f2|...|fw]⊗ h
+∑
(−1)εi+|fi|g ⊗ [f1|...|fifi+1|...|fw]⊗ h
+ (−1)εng ⊗ [f1|...|fw−1]⊗ fwh
where ∂ is the usual tensor product differential on A⊗Tco(ΣA)⊗A, and where εi = (i−1)+|f1|+...+|fi−1|.The A-bilinear map A⊗πBunrA⊗πA
∼−→ A is a quasi-isomorphism.
As a graded vector space the unreduced Hochschild cochain complex C∗unr(A,A) is
Hom(BunrA,A) =∏w≥0
x0,...,xw
Hom(Σ(xwAxw−1)⊗ · · · ⊗ Σ(x1Ax0), xwAx0
).
The differential may be defined through the identification C∗unr(A,A) ∼= HomA−A(A⊗πBunrA⊗πA,A). It
is given by the classical formula of Hochschild: the boundary ∂π(φ) of φ ∈ C∗unr(A,A)n is determined by
setting ∂π(φ)([f1|...|fw]) equal to
∂(φ)([f1|...|fw]) + (−1)|f1|n+1f1φ([f2|...|fw])
+∑
(−1)εi+|fi|φ([f1|...|fifi+1|...|fw])
+ (−1)εnφ([f1|...|fw−1])fw.
And the Hochschild cohomology of A is by definition HH∗(A,A) = H∗(C∗unr(A,A)). If A is augmented
then the canonical restriction C∗unr(A,A)→ C∗(A,A) would be a quasi-isomorphism, so this is consistent
with our previous definitions.
As Keller remarks in [79] if A → A′ is a fully faithful embedding of dg categories then one has
a restriction morphism C∗unr(A′, A′) → C∗unr(A,A) by forgetting those objects which do not come from
A. This fact is visible from the above formulas. In particular there is a canonical homomorphism
HH∗(A′, A′)→ HH∗(A,A). This functoriality will be extremely useful to us later in chapter 4.
2.6 Semi-Free Extensions and Acyclic closures
This section contains background of semi-free dg algebras and semi-free divided power algebras. Much
more detail can be found in [12].
Let A be a strictly graded commutative ring, and let X be a graded set. The free strictly graded
commutative A algebra on X is denoted A[X], see [12, section 2.1] for an explicit description. To
Chapter 2. Background 31
emphasise a different point of view we may also use the notation SymA(X), or Sym(X) when A is
understood.
If A is a dg algebra and z ∈ Ai is a cycle and x is a variable of degree i+ 1, then A[x] has a unique
dg algebra structure for which A → A[x] is a morphism of dg algebras and ∂(x) = z. We may write
A[x] = A[x : ∂(x) = z] to be explicit.
The morphism A → A[x] is an example of a semi-free extension. In general, we define a semi-free
extension to be any morphism A→ A[X] obtained by repeatedly adjoining variables in this way. That
is, X should be a graded set with an increasing filtration ∅ = X(0) ⊆ X(1) ⊆ · · · and A[X] =⋃nA[X(n)]
with ∂(X(n)) ⊆ A[X(n−1)]. In this thesis we work always with positively graded dg algebras, so we can
and will take X(n) = X≤n without comment. In the dual, cohomological situation encountered in rational
homotopy theory this doesn’t work, and the filtration on X should not be neglected. The following lifting
property for semi-free extensions is fundamental. It is proven by induction on the filtration X≤n as in
[12, proposition 2.1.9].
Lemma 5. Let A→ A[X] be a semi-free extension. If φ : B → C is a surjective quasi-isomorphism of
A algebras, then any morphism of A algebras A[X]→ B lifts along φ.
A B
A[X] C
'
That is, a morphism exists as above, making the diagram commute.
Now we assume A is a local dg algebra. A semi-free extension A → A[X] is called minimal if it is
minimal as a homomorphism of local dg algebras, which means ∂(mA[X]) ⊆ mA + m2A[X].
Proposition 4. Any surjection A → B of local dg algebras can be factored as A → A[X] → B, where
A→ A[X] is a minimal semi-free extension and A[X]→ B is a surjective quasi-isomorphism. Any two
such factorisations are isomorphic. The set X is degree-wise finite and concentrated in positive degrees.
Such a factorisation is called a minimal model for the surjection A → B. We will discuss in the
next section how Cohen factorisations are used to deal with the non-surjective case. The first and last
sentences of the proposition can be found in [12, proposition 2.1.10]. By lemma 5 any two minimal models
are connected by a quasi-isomorphism A[X]→ A[X ′] (compatible with A and B). By Nakayama’s lemma
it suffices to show that k[X]→ k[X ′] is an isomorphism, and this is dealt with by [12, lemma 7.2.3].
Divided Powers
Now we discuss divided power structures. Let A be a strictly graded commutative algebra with an ideal
I. A divided power structure on I is a collection of functions defined on even elements of I
x(0) = 1 and x 7→ x(n) : I2i → I2ni for x ≥ 1,
these functions should satisfy the well-known axioms which can be found for instance in [64, section 1.7].
We will call (A, I) a divided power algebra. A homomorphism of divided power algebras φ : (A, I) →(A′, I ′) should satisfy φ(I) ⊆ I ′ and φ(x(n)) = φ(x)(n).
Chapter 2. Background 32
If A contains the field of rational numbers then any proper ideal supports a unique divided power
structure given by x(n) = 1n!x
n.
A key example is the free divided power algebra on a variable u of degree i. Let A be any strictly
graded commutative algebra. If i is odd then A〈u〉 = A[u] is the usual exterior algebra A + uA with
u2 = 0. If i is even then by definition A〈u〉 = A+ uA+ u(2)A+ ... where each u(n) formally generates a
rank one free module with |u(n)| = ni. Multiplication is given by the rule
u(m)u(n) =(m+nm
)u(m+n)
where(m+nm
)= (m+n)!
m!n! . Alternatively one can define Z〈u〉 to be the Z subalgebra of Q[u] generated by
un/n! for n ≥ 0, and then set A〈u〉 = A⊗Z Z〈u〉.More generally, if U is a degree-wise finite non-negatively graded set then by definition A〈U≤n〉 is
the tensor product⊗
AA〈u〉 over u ∈ U≤n, and A〈U〉 =⋃nA〈U≤n〉 along the natural embeddings
A〈U≤n〉 → A〈U≤n+1〉. An alternative notation for A〈U〉 is ΓA(U), or Γ(U) when A is understood.
Throughout this thesis we will use the notation (U)(n) for the ideal in A〈U〉 generated by all mono-
mials u(n1)1 · · ·u(nk)
k with n1 + ...+ nk ≥ n and ui ∈ U .
Lemma 6. The graded algebra A〈U〉 has a unique divided power structure on (U)(1) extending the
functions u 7→ u(n) for u ∈ U .
This is more general than [64, proposition 1.7.6], but the proof there can be adapted to work in this
situation. The free divided power algebra A〈U〉 is characterised by the following important property.
Lemma 7. Let A → B be a homomorphism of strictly graded commutative algebras, and suppose
that B has a divided power structure on an ideal I. If U is a degree-wise finite non-negatively graded
set then any graded function U → I extends uniquely to a homomorphism of divided power algebras
(A〈U〉, (U)(1))→ (B, I)
Again this is more general than [64, lemma 1.7.8] but the proof is similar.
Now we introduce some notation which will be useful later. Let B be a dg algebra with a divided
power structure on a semi-ideal I. The indecomposables of B (with respect to this structure) are denoted
indγ(B) = indγ(B, I) = mB/(m2B + I(2)).
A homomorphism A → A[X]〈U〉 will be called a mixed free extension. We can combine lemma 7
and the usual universal property for free extensions to obtain a universal property for mixed extensions.
It says that homomorphisms from A[X]〈U〉 to an A algebra B with divided powers on I correspond
bijectively to graded functions X → B and U → I.
Lemma 8. Let (A, k) be a local graded algebra with two mixed extensions A[X]〈U〉 and A[Y ]〈V 〉, with
X,U, Y and V all degree-wise finite and in strictly positive degrees. Suppose we have a morphism
φ : A[X]〈U〉 → A[Y ]〈V 〉
of A algebras with divided powers on (U)(1) and (V )(1) respectively. Then φ is an isomorphism if and
only if indγ(φ⊗A k) : indγ(k[X]〈U〉)→ indγ(k[Y ]〈V 〉) is an isomorphism.
Proof. The proof is almost the same as [64, lemma 1.8.7]. It is only necessary to show φ is an isomorphism
assuming that indγ(φ⊗A k) is.
Chapter 2. Background 33
This assumption and the hypotheses imply that X and Y , and U and V respectively, have the same
(graded) cardinality. It follows that each φi : A[X]〈U〉i → A[Y ]〈V 〉i is a map of free A0 modules of equal
finite rank. Since A0 is Noetherian it therefore suffices to show that φ is surjective.
Assume towards a proof by induction that we know φ<i is surjective. Since indγ(φ⊗Ak)i is surjective
0 =A[Y ]〈V 〉i
(mA, im(φ)i, (Y )2, (V )(2))i=
k[Y ]〈V 〉imA0
·A[Y ]〈V 〉i + (A>0 ·A[Y ]〈V 〉<i)i + im(φ)i + (Y )2i + (V )
(2)i
.
By the induction hypothesis (A>0 · A[Y ]〈V 〉<i)i ⊆ im(φ)i and (Y )2i + (V )
(2)i ⊆ im(φ)i. Therefore
A[Y ]〈V 〉i/(mA0·A[Y ]〈V 〉i + im(φ)i) = 0 and by Nakayama’s lemma φi is surjective.
Next we discuss how this structure is used in the dg situation. Let A be a strictly commutative dg
algebra with a semi-ideal I, that is, we assume I is an ideal in the underlying graded algebra of A but
we do not impose the condition ∂(I) ⊆ I. A divided power structure on I is a divided power structure
in the graded sense above which additionally satisfies ∂(x(n)) = ∂(x)x(n−1) for all n ≥ 1.
Let A be dg algebra with divided power structure on I, and let z ∈ Ai be a cycle. If u is a variable of
degree i+1 then A〈u〉 is admits a unique differential making it a dg algebra with divided power structure
on (z)(1) + I, for which A→ A〈u〉 is a morphism of dg algebras with divided powers and ∂(u) = z. We
may write A〈u〉 = A〈u : ∂(x) = z〉 to be explicit.
The morphism A → A〈u〉 is an example of a Tate extension, or a semi-free divided power extension
(often also called a semi-free Γ extension). In general a Tate extension is any morphism A → A〈U〉obtained as a colimit of such extentions, so that A〈U〉 =
⋃nA〈u1, ..., un〉 with ∂(un) ∈ A〈u1, ..., un−1〉,
where U = u1, u2, ... is a graded set of variables.
Then analogous lifting property for Tate extensions is similar to lemma 5 (with almost the same
proof), but delicate to state. We will not need it below, so we skip it (but note that we will make the
same lifting argument by hand to prove theorem 21).
A Tate extension A → A〈U〉 is called minimal if ∂(mA〈U〉) ⊆ mA + (U)(2). Analogous to lemma 4,
every surjection of local dg algebras A → B can be factored as a minimal Tate extension A → A〈U〉followed by a surjective quasi-isomorphism A〈U〉 → B. This factorisation is call a minimal Tate model
or an acyclic closure of the homomorphism A→ B.
The Tate Model for the Diagonal
We end this section by discussing the classical Tate model for the diagonal. Inspired by rational homotopy
theory, Tate used semi-free divided power extensions to build dg algebra resolutions [117]. In particular,
Tate wrote down the minimal resolution of the residue field of a complete intersection local ring (see [64,
proposition 1.5.4]). The theorem below is a structural version of his construction which summarises and
slightly extends several results from the literature. Similar statements can be found in [44, proposition
1.9], [19, theorem 1.1] and [12, proposition 7.2.9]. Because the statement here is more general than can
be found in the literature, we will sketch a proof. In section 3.4 we will explain how this construction
can be made functorial.
We say that Q is a smooth K algebra if we are given a homomorphism of commutative Noetherian
rings K → Q which is essentially of finite type, and such that the Kernel I of the multiplication map
µ : Q ⊗K Q → Q is locally generated by a regular sequence. This means that Ip is generated by a
Chapter 2. Background 34
regular sequence for every prime p of the enveloping algebra Qev = Q⊗KQ. See for example [20, section
1] for discussion and other definitions of smoothness. In particular if Q is local with maximal ideal m
and n = µ−1(m) then In is generated by a regular sequence x. We continue to use this notation in the
theorem below. Note that Qev is Noetherian since K → Q is essentially of finite type.
We will often apply the theorem in the situation K = Q, then smoothness is trivial and our notation
becomes less complicated (we will make this simplifying assumption in section 3.4). Another important
case is when K is a local ring and Q is K[x1, ..., xn] localised at (mK , x1, ..., xn).
If Q→ Q[X] is a semi-free extension, the double Q[X]⊗K Q[X] is simply denoted Qev[X,X] or even
Q[X]ev. For geometric reasons, the multiplication map Qev[X,X]→ Q[X] is referred to as the diagonal.
Theorem 10. Let (Q,m, k) be a local ring which is also a smooth K algebra as above, and let Q→ Q[X]
be a semi-free extension with X degree-wise finite and concentrated in strictly positive degrees. The
multiplication map Qevn [X,X]→ Q[X] has a Tate model of the form
Qevn [X,X]→ Qev
n [X,X]〈Σx, ΣX〉 → Q[X].
The differential satisfies ∂(Σx) = x for x ∈ x and ∂(Σx) = 1⊗ x− x⊗ 1− Σ∂1(x) modulo decomposables
in m2Q[X]evn 〈Σx,ΣX〉
+ (ΣX)(2) for x ∈ X.
If moreover Q → Q[X] is minimal as a semi-free extension then Qevn [X,X]〈Σx, ΣX〉 is minimal as a
dg module over Qevn [X,X], that is ∂(ΣX) ⊆ (mQ[X]evn
)Q[X,X]〈ΣX〉.
In the statement of the theorem ∂1 : XQ → XQ is the strict part of the differential. That is, the
part which preserves the weight decomposition Q[X] =⊕
w SymwQ(X). The ∂1(x) in the statement is
really some lift of ∂1(x) to XQev. Any lift is equally good, and we continue with this abuse of notation
below.
Lemma 9. Any map Qevn [X,X]〈Σx, ΣX〉 → Q[X] whose differential has the form described in the theorem
above is a quasi-isomorphism.
Because x is a regular sequence there is an obvious quasi-isomorphismQevn [X,X]〈Σx, ΣX〉 → Q[X,X]〈ΣX〉.
Then by a homological version of Nakayama’s lemma it suffices to apply − ⊗LQ k and show that
k[X,X]〈ΣX〉 → k[X] is a quasi-isomorphism. Finally this is essentially contained in [44, proposition
1.9] (the proof is an application of lemma 1).
Proof of theorem 10. Assume inductively that we have constructed a factorisation
Qevn [X,X]→ Qev
n [X,X]〈Σx,X<n〉 → Q[X]
whose differential has the above form. Given x ∈ Xn we must define ∂(Σx). Automatically the differential
restricts to Qevn [X<n, X<n]〈Σx, ΣX<n〉. Consider ∂[1 ⊗ x − x ⊗ 1 − Σ∂1(x)], by the induction hypothesis
this a cycle in the kernel I → Qevn [X<n, X<n]〈Σx, ΣX<n〉 → Q[X<n]. By the previous lemma I is acyclic,
so there is y ∈ In such that ∂(y) = ∂[1⊗ x− x⊗ 1− Σ∂1(x)]. Automatically, y is decomposable, because
otherwise for degree reasons we could write y = aΣx′+p with p a decomposable in m2Q[X]evn 〈Σx,ΣX〉
+(ΣX)(2)
and x′ ∈ Xn−1 and a ∈ Qevn rnQev
n . But then by the induction hypothesis we would have ∂(y) = a⊗x′−x′⊗a−aΣ∂1(x′)+q with q also decomposable. This is not the case since ∂[1⊗x−x⊗1− Σ∂1(x)] involves
no terms of the form aΣ∂1(x′), using the fact that ∂21⊗Qk = 0. So setting ∂(Σx) = 1⊗x−x⊗1−Σ∂1(x)−y
finishes the induction. Taking the colimit builds the Tate model for the diagonal.
Chapter 2. Background 35
For the final statement, first note that minimality of Q → Q[X] implies ∂1 ⊗Q k = 0, and therefore
by the description of the differential Qevn [X,X] → Qev
n [X,X]〈Σx, ΣX〉 is a minimal Tate extension. In
other words, Qevn [X,X]→ Qev
n [X,X]〈Σx, ΣX〉 → Q[X] is an acyclic closure.
Applying −⊗Qevn [X]〈Σx〉 k produces an acyclic closure k[X]〈ΣX〉 → k. It suffices to show that this is a
minimal dg module resolution, that is, ∂(ΣX) ⊆ (X)k[X]〈ΣX〉. This is [12, theorem 6.3.4].
Remark 1. For simplicity, assume K = Q. One can also construct the Tate model for the diagonal
Q[X,X]〈ΣX〉 through Koszul duality methods. In short Q[X] = CQL is the cobar construction of a curved
Lie∞ coalgebra L over Q (see section 2.8 for some details in this direction). Its universal envelope UL is
a curved A∞ coalgebra, and there is an acyclic twisting cochain τ : UL→ CQL over Q. One can use τ to
build a twisted tensor product and an isomorphism CQL⊗τQ UL⊗τQ CQL ∼= Q[X,X]〈ΣX〉. In particular,
the appearance of divided powers here can be explained by the Poincare-Birkhoff-Witt theorem. Some
of these words are explained (in much less generality) in section 2.8. See especially the proof of theorem
13. This construction will be explained properly somewhere else. This is also connected to joint work
with Vincent Gelinas on explicit universal envelopes of Lie∞ algebras, currently in preparation [33].
The case of a Koszul algebra over a field (when there is no higher structure) is a relatively easy
verification. This is essentially lemma 13 below.
Example 3. The situation for complete intersections is classical. Let K be a commutative ring and set
Q = K[x1, ..., xn] (we won’t bother to localise as this wasn’t important in theorem 10). The kernel of
µ : Q⊗KQ→ Q is generated by the regular sequence x = xi⊗1−1⊗xi. Now take a regular sequence
f1, ..., fc in Q and set R = Q/(f1, ..., fc). As above we write Qev = Q⊗K Q and Rev = R⊗K R.
The Koszul complex Q → Q[η]'−→ R is a semi-free model for R, where the η = η1, ...ηc are degree
one variables with ∂(ηi) = fi.
According to theorem 10 we can build an acyclic closure Qev[η, η] → Qev[η, η]〈Σx, Ση〉 '−→ Q[η]. Ten-
soring down to R we obtain an acyclic closure Rev → Rev〈Σx, Ση〉 '−→ R. This is the minimal Rev free
resolution of R. This construction goes back to Tate (he used it to resolve the residue field of R, see [64,
proposition 1.5.4]).
Guccione and Guccione [60] use a similar construction to compute the Hochschild homology and
cohomology of R over K. Assume that K → R is flat (otherwise we compute Shukla homology instead).
In short, HH∗(R/K,R) is the homology of
(Rev〈Σx, Ση〉)⊗Rev R ∼= R〈Σx, Ση〉.
In characteristic zero this appears already in the work of Wolffhardt [122], and implicitly in the work
of Quillen [106]. Note that in this dg algebra ∂(x) = 0. From this we see that R〈Σx, Ση〉, and hence
HH∗(R/K,R), splits into a direct sum of components R〈Σx, Ση〉(w) =∑i+j=w ΓiR(Ση) ⊗R
∧jR(Σx). This
is the Hodge decomposition established in [60].
Similarly if R is projective over K then HH∗(R/K,R) is the homology of HomRev(Rev〈Σx, Ση〉, R) ∼=R[y, ξ] where y = (Σx)∨ and ξ = (Ση)∨. Hochschild cohomology also obtains a Hodge decomposition by
these considerations.
The same argument gives a short proof of the more general result in [61]: all that was required was
for the map K → Q to be smooth.
Finally, note that we obtain resolutions M ⊗K R〈Σx, Ση〉 '−→M for any R module M . In particular if
K = k is the residue field of R (so that R is an honest complete intersection ring) we have a resolution
Chapter 2. Background 36
R〈Σx, Ση〉 '−→ k, but this time R〈Σx, Ση〉 has a different differential (it can be written as a double complex
in such a way that the vertical part computes Hochschild homology).
Remark 2. Theorem 10 can be used to compute Hochschild homology (and cohomology) much more
generally. Let Q be a local ring which is a smooth K algebra and let R = Q/I be a factor ring of Q.
Assume that K → R is flat (otherwise we compute Shukla homology instead). Denote Rev = R ⊗K R
and Qev = Q⊗K Q, with n and x as above as well.
The theorem allows us to compute HH∗(R/K,R) from a minimal model Q → Q[X]'−→ R. We first
need the observation that there is a canonical quasi-isomorphism R ⊗LRev R ' R ⊗L
RevnR, because R is
already n-local. From the resolution Qevn [X,X]〈Σx, ΣX〉 → Q[X] we obtain a resolution Rev
n 〈Σx, ΣX〉 → R.
Hence R⊗LRev
nR ' (Rev
n 〈Σx, ΣX〉)⊗RevnR ∼= R〈Σx, ΣX〉.
Therefore HH∗(R/K,R) is the homology of the semi-free divided power algebra R〈Σx, ΣX〉. Similarly,
we can produce a dual model for HH∗(R/K,R) using theorem 10.
Note that in the cases Q = R we recover the famous Hochschild-Kostant-Rosenberg theorem.
Remark 3. The Tate model for the diagonal also easily reproduces a result of Fiorenza and Manetti
on mapping cones of Lie∞ algebras [50]. This will be explained in section 3.5 (see corollary 5).
2.7 Cohen Factorisations and Minimal Models
In this section we present the basic facts we need about the theory of Cohen factorisations, introduced
by Avramov, Foxby and Herzog [16]. We will also introduce an important class of dg algebras which we
call minimal models.
Let φ : (R,m, k) → (S, n, l) be a local homomorphism. By definition φ is weakly regular if it is
flat and the fibre S/mS is a regular ring. A regular factorisation of φ is a factorisation of φ into local
homomorphisms
R′
R Sφ
such that R→ R′ is weakly regular and R′ → S is surjective.
For example, φ is said to be essentially of finite type if for some finite sequence x = x1, ..., xn of
variables φ can be factored as an extension R → R[x](x) followed by a surjection R[x](x) → S. Largely,
these are the homomorphisms of interest in algebraic geometry. Since R → R[x](x) is weakly regular,
such a homomorphism admits a regular factorisation by definition.
A regular factorisation R → R′ → S in which R′ is complete is called a Cohen factorisation. It is
an important theorem of Avramov, Foxby and Herzog that any local homomorphism R→ S in which S
is complete admits a Cohen factorisation [16, (1.1)]. By this means, these authors initiated a program
for investigating the local properties of ring homomorphisms, using Cohen factorisations to reduce to
properties of surjective homomorphisms.
The composition R → S → S of φ with the natural embedding of S into its completion is known
as the semi-completion of φ, and is generally denoted φ. By Avramov, Foxby and Herzog’s theorem
φ : R→ S always admits a Cohen factorisation.
A morphism between two Cohen factorisations R → R′ → S and R → R′′ → S is a commutative
Chapter 2. Background 37
diagram
R′
R S
R′′
ψ
of local homomorphisms. The factorisation R → R′ → S is said to be a deformation of R → R′′ → S
(and R → R′′ → S a reduction of R → R′ → S) if ψ is surjective. In this situation the kernel of ψ is
automatically generated by a regular sequence x. Moreover, if I denotes the kernel of R′ → S, then x
descends to an l vector space basis for (I+mR′+m2R′)/(mR
′+m2R′). Conversely, any lift x ⊆ I of a basis
for (I + mR′ + m2R′)/(mR
′ + m2R′) is a regular sequence in R′ and reducing modulo x produces a Cohen
factorisation R → R′′ = R′/(x)→ S which is a reduction of R → R′ → S. These facts can be found in
[16, (1.5)], where it is also observed that diml(I+mR′+m2R′)/(mR
′+m2R′) = edimR′/mR′−edimS/mS.
With this in mind a Cohen factorisation R → R′ → S of φ is said to be minimal if edimR′/mR′ =
edimS/mS, or equivalently, if it admits no proper reduction.
While Cohen factorisations cannot necessarily be compared directly, by [16, (1.2)] any two Cohen
factorisations R→ R′ → S and R→ R′′ → S admit a common deformation
R′
R A S
R′′.
So, the horizontal row is a Cohen factorisation and the two vertical homomorphisms are surjective. More-
over, any two Cohen factorisations of φ can be connected by a sequence of codimension one deformations.
This observation is often useful to simplify arguments which compare Cohen factorisations.
Complete Intersections and Quasi-Complete Intersection
Now we discuss some classes of homomorphisms which have been introduced using Cohen factorisations.
A surjective homomorphism is called complete intersection if its kernel is generated by a regular
sequence. In general a local homomorphism R→ S is complete intersection if in some Cohen factorisation
R → R′ → S the homomorphism R′ → S is complete intersection. This definition was introduced by
Avramov in [13], where it was used to settle some long-standing conjectures of Quillen.
It follows readily from the definition that any complete intersection homomorphism has finite flat
dimension. The next definition is an analogue which allows for infinite flat dimension.
A surjective local homomorphism R→ S is called a quasi-complete intersection (qci) if it admits an
acyclic closure R → R〈U〉 '−→ S in which U is concentrated in degrees 1 and 2. These homomorphisms
were introduced by Blanco, Majadas and Rodicio in [31], where they were characterised in terms of
vanishing of Andre-Quillen cohomology. In [18, section 7] this class is extended to include non-surjective
homomorphisms: R→ S is qci if in some Cohen factorisation R→ R′ → S the homomorphism R′ → S
is qci.
In both cases the property is independent of the choice of Cohen factorisation (the arguments can
be found in the references given).
Chapter 2. Background 38
Minimal Models
Now we explain the construction of minimal models, which will be extremely important throughout this
thesis.
Let R be a local ring. The classical Cohen structure theorem is the absolute analogue of the existence
of Cohen factorisations. It says that there is a complete, regular local ring Q and a surjection Q →R. We can and will assume this presentation is minimal in the sense that ker(Q → R) ⊆ m2
Q. By
proposition 4 there is a minimal model Q→ Q[X]→ R. Note that Q[X] is in fact absolutely minimal,
so ∂(mQ[X]) ⊆ mQ[X]. The dg algebra Q[X] will be called a minimal Cohen model for R.
We can use the theory of Cohen factorisations to make a relative version of this construction. Let
φ : (R, k) → (S, l) be a local homomorphism. The semi-completion of φ admits a minimal Cohen
factorisation R → R′ → S. By proposition 4 there is a minimal model R′ → R′[X] → S. The
factorisation R → R′[X] → S will be called a minimal Cohen model for φ. If Q = k ⊗R R′, then the
fibre k⊗RR′[X] = Q[X] is a dg algebra model for the completion of the derived fibre k⊗LR S. Note that
k → Q[X] is a minimal dg algebra morphism, and therefore Q[X] is absolutely minimal.
In both situations Q[X] is a semi-free extension of a regular local ring which is also absolutely
minimal. In an abuse of terminology, we will call a local dg algebra of this form a minimal model. These
dg algebras will be very important in what follows.
In section 2.4 minimal A∞ algebras were also called minimal models. Needless to say these are two
very separate (but related) notions. The definition given here will be used exclusively in chapter 3, while
the term minimal model is reserved for minimal A∞ algebras in chapter 4.
Golod Homomorphims
Another class of homomorphisms for which Cohen factorisations are elucidating are Golod homomor-
phisms. However, these do not fit into the same basic pattern as complete intersection and quasi-complete
intersection homomorphisms.
Let φ : (R,m, k)→ (S, n, l) be a finite local homomorphism. We recall the classical bound
PSl (t) ≤ PRk (t)
(1 + t)− tPRS (t)
which is a coefficient-wise inequality of formal power series. The bound appears in [57] where it is
attributed to Serre. In loc. cit. Golod characterised those surjective ring homomorphisms for which this
bound is attained in terms of trivial Massey operations on TorR∗ (S, k).
Classically φ has been called Golod if the Serre bound above is an equality. However, in certain
exceptional cases this bound can be improved and Avramov gives a definition in [10] which takes this
into account. We give an equivalent definition now.
Let R → R〈U〉 '−→ k be an acyclic closure, so that S〈U〉 ' S ⊗LR k. Avramov defines φ to be Golod
if S〈U〉 is a Golod dg algebra in the sense of [10, definition (2.2)]. Take also a minimal Cohen model
R → R′[X]'−→ S and set A = k ⊗R R′[X]. It is easy to see that the Golod condition for dg algebras
is invariant under completion and quasi-isomorphisms, therefore A is Golod if and only if S〈U〉 is so.
Finally, it is most convenient to use the characterisation [10, theorem 3.4 (3)] for our definition: we say
that φ is Golod if the map mA → mA/m2A induces an injection H∗(mA) → H∗(mA/m
2A) = mA/m
2A. The
relation to the homotopy Lie algebra will be discussed in example 4.
Chapter 2. Background 39
Note that this definition implies that the fibre S = S/mS is Artinian with m2S
= 0. In particular
R → S is automatically module finite. To account for the non-finite case one might want to extend
the definition using Cohen factorisations. The next proposition suggests such a generalisation. We
present it only as an example of how Cohen factorisations are used to introduce new properties of ring
homomorphisms.
Let us emphasise that the proposition is for the most part a consequence of the results of Avramov
in [10], which builds on work of Golod [57] and Levin [87].
Following [10, (4.4)], a surjective homomorphism (R,m)→ (S, n) is called exceptional if its kernel is
generated by a single non zero-divisor in mrm2, otherwise it is standard. Exceptional homomorphisms
are Golod, but as the name suggests they are not representative of the typical situation.
Proposition 5. If φ : (R,m, k)→ (S, n, l) is a local homomorphism then the following are equivalent:
1. for some Cohen factorisation R→ R′ → S the equality
PSl (t) =PR
′
l (t)
(1 + t)d+1 − tPR′S
(t)
holds, where d = edim(R′/mR′)− edim(S/mS);
2. the above equality holds for any Cohen factorisation;
3. in some minimal Cohen factorisation R→ R′ → S the map R′ → S is standard and Golod;
4. in any minimal Cohen factorisation R→ R′ → S the map R′ → S is standard and Golod.
The proposition will not be important to us later, and it is partly superseded by proposition 9, so
we sketch the proof only. Proposition 9 will characterise the equivalent conditions here in terms of L.S.
category. With that result in mind, a homomorphism satisfying the conditions above will be called
generalised Golod of level 1.
We will need a standard lemma.
Lemma 10. If (R,m, k) is a local ring and x is a non zero divisor in mrm2 then PRM (t) = (1+t)PR/xM (t)
for any R module M such that xM = 0.
The proof uses the simple calculation M ⊗LR k 'M ⊗L
R/x k + ΣM ⊗LR/x k.
Proof of proposition 5. Since d = 0 if and only if the Cohen factorisation is minimal, the implications
2 =⇒ 4 =⇒ 3 =⇒ 1 are clear using the characterisation [10, theorem (4.6)]. Thus we must show
that show that 1 =⇒ 2.
For this we appeal to the fact that any two Cohen factorisations R → R′ → S and R → R′′ → S
are connected by a sequence of deformations. So we can assume we have a homomorphism R′′ → R′
(compatible with R and S) whose kernel is generated by a nonzero divisor x in mR′′ r (m2R′′ + mR′′),
and we must show that the formula in part 1 holds for R′ if and only if it holds for R′′.
Let e = edim(R′′/mR′′)−edim(S/mS) and d = edim(R′/mR′)−edim(S/mS), and note that e = d+1
since edim(R′′/mR′′) = edim(R′/mR′) + 1. Now we use lemma 10 to make the calculation
PR′′
l (t)
(1 + t)e+1 − tPR′′S
(t)=
(1 + t)PR′
l (t)
(1 + t)e+1 − (1 + t)tPR′
S(t)
=PR
′
l (t)
(1 + t)d+1 − tPR′S
(t).
Chapter 2. Background 40
This concludes the proof.
Note that in the exceptional cases we have the improved bound
PSl (t) =PR
′
l (t)
(1 + t)d+2 − tPR′S
(t).
Remark 4. The Serre bound above is attained by the famous bar resolution. Details on this construction
and the A∞ structure which underlies it are given by Burke in [39].
Let Q → R be a standard, surjective, Golod homomorphism, let A → R be the minimal Q free
resolution of R, and let K → k be the minimal Q free resolution of k. Burke constructs an A∞ algebra
structure on A making A → R a quasi-isomorphism of A∞ algebras, and an A∞ A module structure
on K compatible with the R module structure of k. From this data one constructs the twisted tensor
product K ⊗Q BA ⊗Q R, and Burke shows that the Golod property implies that K ⊗Q BA⊗Q R → k
is a minimal R free resolution.
Now suppose that R → S is generalised Golod of level 1. Let R → R′ → S be a minimal Cohen
factorisation, let A → S be the minimal R′ free resolution of S, and let K → k be the minimal R′
free resolution of k, note that these are both R flat resolutions. Then R′ → S is Golod so using the
constructions of Burke we obtain a minimal S free resolution K ⊗R′ ⊗BA⊗R′ S → k.
Gulliksen and Schoeller’s Theorem
To end this section we record the the following important consequence of theorem 10. It was proven
independently by Gulliksen and Schoeller.
Theorem 11 (Gulliksen [64], Schoeller [109]). If Q[X] is a minimal model with residue field k and X0
is a minimal generating set for mQ then we have an acyclic closure Q[X]〈t, ΣX〉 '−→ k where t is a set of
variables in degree 1 with ∂(t) = X0. Moreover Q[X]〈t, ΣX〉 is minimal as a dg module over Q[X].
In particular if Q[X]→ R is a minimal Cohen model we have an acyclic closure R〈t, ΣX〉 '−→ k which
is minimal as a complex. Even if R does not admit a minimal Cohen model (without completing), the
acyclic closure of R→ k is minimal as a complex.
If Q→ Q〈t〉 '−→ k is the Koszul complex which resolves k over Q, then lemma 5 produces a Q algebra
morphism Q[X] → Q〈t〉. We obtain a quasi-isomorphism Q[X]〈t, ΣX〉 = Q[X,X]〈ΣX〉 ⊗Q[X] Q〈t〉'−→
Q〈t〉 '−→ k. When Q[X] is a minimal Cohen model for R the quasi-isomorphism R〈t, ΣX〉 '−→ k is given
by adjunction. The last statement follows using the faithfully flat base change −⊗R R and uniqueness
of acyclic closures.
2.8 Lie Algebras and Lie Coalgebras
In this section we will quickly define graded Lie algebras and graded Lie coalgebras, and explain why
both of these objects are equivalent to certain semi-free dg algebras. Everything in this section takes
place over a fixed field k, about which we make no additional assumptions. The content of this section
is classical.
Chapter 2. Background 41
We define graded Lie algebras in terms of operations (bracket and reduced square) following [12,
section 10], where more details can be found. This approach leads to a complicated definition, but one
which should be at least partially familiar to all readers.
A graded Lie algebra is a graded vector space L over k equipped with a bilinear bracket [ , ] :
L⊗ L → L and a reduced square defined on odd elements ( )[2] : L2i+1 → L4i+2 which is quadratic, so
(αu)[2] = α2u[2] for any odd u in L and α in k. These operations should satisfy the following axioms.
• Compatibility: for any odd elements u, v in L
(u+ v)[2] = u[2] + [u, v] + v[2].
• Anti-symmetry: for any x, y, z in L with z even
[x, y] + (−1)|x||y|[y, x] = 0 and [z, z] = 0.
• Jacobi identity: for any x, y, z, u, v in L with u odd
[x, [y, z]] = [[x, y], z] + (−1)|x||y|[y, [x, z]], [u[2], v] = [u, [u, v]] and [u, [u, u]] = 0.
For example, any graded associative algebra A over k gives rise to a graded Lie algebra ALie with the same
underlying graded vector space. The bracket is given by the graded commutator [x, y] = xy−(−1)|x||y|yx
while the reduced square is simply u[2] = u2 for any x, y, u in A with u odd.
A morphism of graded Lie algebras is a map of graded vector spaces φ : L → L′ which respects the
two operations, so that φ[x, y] = [φ(x), φ(y)] and φ(u[2]) = (φ(u))[2] for any x, y, u in L with u odd.
The category of graded Lie algebras over k is denoted Liegr, leaving k implicit. We further assume
that the objects of Liegr are degree-wise finite dimensional and concentrated in strictly positive degrees.
Gradings here will be written cohomologically.
In characteristic zero a Lie coalgebra may be defined as a vector space L with a cobracket L→ L⊗Lwhich is anti-symmetric and satisfies the Jacobi condition laid out for instance in [108, section 1]. In
general, the symmetry condition is more subtle and it is easiest to define Lie coalgebras as follows.
First, if V is a graded vector space then the exterior algebra∧∗
(V ) is the quotient of the tensor
algebra T(V ) by the two sided ideal(u⊗ v + (−1)|u||v|v ⊗ u, w ⊗ w : u, v, w ∈ V with w even
). Then∧∗
(V ) has an internal grading inherited from V and an additional weight grading, with∧w
(V ) =
(v1 · · · vw : vi ∈ V ) spanned as a vector space by products of length w.
A graded Lie coalgebra over k is firstly a graded vector space L together with a cobracket ∆ : L →∧2L. This gives rise to a map ∆ ∧ 1 − 1 ∧ ∆ :
∧2L →
∧3L, and we impose the condition that
(∆ ∧ 1 − 1 ∧∆)∆ = 0. How this relates to the Jacobi identity is explained in the proof of theorem 12
below (which we sketch only).
A Lie∞ coalgebra is the structure one gets by allowing the cobracket to take arbitrary positive weight
∆ = ∆1 + ∆2 + ∆3 + ... : L→∧>0
L. For details in characteristic zero see [84] or [90], and in arbitrary
characteristic there is some discussion in [28].
Any graded associative coalgebra C has an associated Lie coalgebra CLie with the same underlying
graded vector space. The cobracket is simply the coproduct C → C ⊗ C composed with the natural
projection C ⊗ C →∧2
C.
Chapter 2. Background 42
A morphism of Lie coalgebras is a map of graded vector spaces φ : L → L′ making the obvious
diagram commute:
L∧2L
L′∧2L′.
∆
φ∧2 φ
∆
By definition coLiegr is the category of degree-wise finite dimensional graded Lie coalgebras which are
concentrated in strictly positive degrees. Gradings here will be written homologically.
Finally, we introduce an important category of quadratic semi-free algebras over k. Once again k
will be left implicit in our notation. The objects of q-semi-free are bigraded algebras of the form
A = k[V ] = Sym(V ) where V is a degree-wise finite dimensional graded vector space with V = V≥0.
They should be graded cohomologically by weight Aw = Symw(V ) and homologically by the internal
degree Ai = k[V ]i = Sym(V )i induced from V , so that A is strictly graded commutative with respect
to its lower grading. The differential must satisfy ∂ : Awi → Aw+1i−1 . Morphisms in q-semi-free are local
homomorphisms of dg algebras which respect both gradings6.
These dg algebras are called quadratic because the differential applied to the generators V takes
values in the quadratics forms Sym2(V ). The relationship between this category and the category of
minimal models will be made clear in section 3.1.
The following theorem is fundamental in the Koszul duality between Lie algebras and commutative
algebras. It is essentially due to Quillen [105, appendix B] in characteristic zero and Avramov [9, theorem
4.2] in positive characteristic.
Theorem 12. We have equivalences of categories
q-semi-free coLiegr Liegr.π∗
C
(−)∨
(−)∨
The proof is a series of verifications, we sketch it only. The functors themselves will be defined as
we go. The key point for the two functors on the left is that there is a natural isomorphism∧w
(ΣV ) ∼=ΣwSymw(V ), where we use Σ to denote the usual suspension acting on the internal degree only.
If A is an object of q-semi-free we define π∗(A) = ΣA1. Writing A = Sym(V ), the graded Lie
coalgebra structure on ΣA1 = ΣV is defined by commutativity of the left square in the following diagram
Vi−1 Sym2(V )i−2 Sym3(V )i−3
(ΣV )i∧2
(ΣV )i∧3
(ΣV )i.
∂
∼=
∂
∼= ∼=
∆ ∆∧1−1∧∆
Since ∂2 = 0, the condition (∆ ∧ 1 − 1 ∧ ∆)∆ = 0 follows after checking that the right square above
commutes.
This process is reversible: if L is an object of coLiegr then CL = Sym(Σ−1L) is a quadratic semi-free
dg algebra with differential defined on V = Σ−1L by the left square in the above diagram. Then ∂
6These correspond to strict morphisms of graded Lie algebras. Working with non-strict morphisms, which may increasethe weight filtration, gives extra flexibility which can be useful. Non-strict morphisms are homotopically no more generalthan strict morphisms, and we won’t need them, so we refer to [90] for more information.
Chapter 2. Background 43
extends uniquely to a derivation on CL, and the right square above implies ∂2 = 0.
The remaining functors in the theorem are just given by taking the k linear graded dual.
Let L be a graded Lie coalgebra. The projection T(L)→∧
(L) is dual to an embedding ι : (∧L)∨ →
T(L∨), and in weight 2 the image may be described as the linear span
ι(∧2
L)∨ =(x⊗ y − (−1)|x||y|y ⊗ x, u⊗ u : x, y, u ∈ L∨ with u odd
).
Using this fact, one may define the operations on the dual L∨ as follows. The cobracket ∆ : L→∧2
L
dualises to a linear map ∆∗ : ι(∧2
L)∨∼=−→ (
∧2L)∨
∆∨−−→ L∨ and we set
[x, y] = ∆∗(x⊗ y − (−1)|x||y|y ⊗ x) and u[2] = ∆∗(u⊗ u).
One could also phrase this in terms of the divided power algebra structure on (Sym(Σ−1L))∨ ∼= Γ(ΣL∨).
The compatibility and anti-symmetry conditions are quite clear already on the level of T2(L∨), because
these identities hold for the tensors x⊗y−(−1)|x||y|y⊗x and u⊗u. The Jacobi conditions are equivalent
to (∆ ∧ 1− 1 ∧∆)∆ = 0. To see this one works in ι(∧3
L)∨, which is similarly spanned by certain anti-
symmetric tensors in L∨ (of three types, corresponding to the three Jacobi identities). We can sketch
the classical Jacobi identity for x, y, x in L∨. Denote the anti-symmetrisation of a tensor x1 ⊗ · · · ⊗ xnby x1 ⊗ · · · ⊗ xn, so for instance x⊗ y = x⊗ y − (−1)|x||y|y ⊗ x. The key identity is
x⊗ y ⊗ z = x⊗ y ⊗ z + (−1)|z|(|x|+|y|)z ⊗ x⊗ y + (−1)|x|(|y|+|z|)y ⊗ z ⊗ x.
With this, one can check easily that
(x⊗ y ⊗ z)ι(∆ ∧ 1− 1 ∧∆)∆ = [[x, y], z] + (−1)|z|(|x|+|y|)[[z, x], y] + (−1)|x|(|y|+|z|)[[y, z], x]
and this vanishes since (∆ ∧ 1 − 1 ∧∆)∆ = 0. This establishes the classical Jacobi identity, the other
two are similar. Having checked this, it follows that a Lie coalgebra structure on L is equivalent to a Lie
algebra structure on L∨, and this finishes the proof of the theorem.
Remark 5. If L is a graded Lie algebra then the composition C(L∨) is the well-known Chevalley-
Eilenberg dg algebra of L.
Remark 6. At the same time, the functor C : coLiegr → q-semi-free is an example of a cobar
construction in the sense of [90]. It can be extended to the category of dg Lie coalgebras (and ultimately,
Lie∞ coalgebras). There is a corresponding bar construction L : Comdg → coLiedg which takes a
strictly commutative augmented dg algebra A to the cofree Lie coalgebra coLie(ΣA) on the suspension
of the augmentation ideal A = ker(A → k). Just as for CL, the Lie coalgebra LA is equipped with a
differential which encodes exactly the product on A. Good properties of the adjunction (C, L) underlie
the Koszul duality between commutative and Lie algebras, see [105] or [70] in characteristic zero. In any
characteristic there is a natural quasi-isomorphism CLA → A, see [51]. On the other hand, L → LCLneed not be a quasi-isomorphism outside of characteristic zero, but it is a weak-equivalence in a certain
sense (unfortunately, not the fully-fledged model category sense). This point is delicate, and we choose to
side-step these difficulties by working almost exclusively with minimal dg algebras, which means focusing
our attention on the left adjoint C.
Chapter 2. Background 44
At this point we will very quickly discuss universal envelopes of Lie algebras and coalgebras. In short,
the two functors A 7→ ALie and C 7→ CLie have left and right adjoints respectively.
In terms of explicit universal properties this means the following. Let L be a graded Lie algebra.
There is a graded associative algebra UL equipped with a map of Lie algebras L → (UL)Lie, and for
any graded associative algebra A, every morphism L→ ALie factors as L→ (UL)Lie → ALie for a unique
algebra map UL → A. The algebra UL is the universal enveloping algebra of L. More details can be
found in [12, section 10].
Similarly, every graded Lie coalgebra L has a universal enveloping coalgebra, which we will denote
in the same way as UL. It enjoys the dual universal property: UL comes with a map of Lie coalgebras
(UL)Lie → L, and for any graded associative coalgebra C, every morphism CLie → L factors as CLie →(UL)Lie → L for a unique coalgebra map C → UL.
Remark 7. Contemplating the construction in the proof of theorem 12 one concludes that the functors
A 7→ ALie and C 7→ CLie are linearly dual to each other, and therefore the two universal properties above
are linearly dual as well. It follows that for any graded Lie coalgebra L in coLiegr there is a canonical
isomorphism (UL)∨ ∼= U(L∨).
As we will see in the proof below, it will actually be better for us to work with the completed cobar
construction CL =∏w CwL. Note that this is the graded product: it only makes a difference if L is
non-zero in degree 1. These completed cobar constructions could be added to the equivalence of theorem
12.
Theorem 13. If L is a graded Lie coalgebra in coLiegr then the composition τ : UL→ L → ΣCL is an
acyclic twisting cochain.
Proof. To check that τ is a twisting cochain we only need the fact that (UL)Lie → L is a Lie coalgebra
morphism. This makes the upper left part of the diagram below commute
UL⊗ UL L⊗ L Σ2CL⊗ CL
UL Σ2CL.∧2
L
L Σ2C2L
ΣC1L
−∼=∆
∼= ∂
The lower square commutes by definition and the upper right part anti-commutes. Comparing the upper
and lower paths yields the equality τ ^ τ + ∂(τ) = 0.
To see that τ is acyclic we verify that the bimodule homomorphism CL ⊗τ UL ⊗τ CL → CL is a
quasi-isomorphism. This is easily seen by reduction to the abelian case, as follows.
Firstly, UL admits a decreasing filtration U (w)L =⋂v<w ker(UL → UL⊗v → L⊗v). Since L is
concentrated in strictly positive degrees this filtration is complete. We also filter CL by weight C(w)L =∏v≥w CvL. The effect of passing to the associated graded vector spaces is to replace L with the abelian
Lie coalgebra Lab on the same space but with cobracket equal to zero. In other words gr(CL) ∼= CLab ∼=k[Σ−1L] and by the Poincare-Birkhoff-Witt theorem gr(UL) ∼= ULab ∼= k〈L〉. We also give CL⊗τUL⊗τ CLthe total weight filtration. By lemma 1 it suffices to show that gr(CL⊗τ UL⊗τ CL)→ gr(CL) is a quasi-
Chapter 2. Background 45
isomorphism. This is exactly the Tate resolution of the diagonal
k[Σ−1L, Σ−1L]〈L〉 → k[Σ−1L] with ∂(x) = Σ−1x⊗ 1− 1⊗ Σ
−1x for x ∈ L,
which is well-known to be a quasi-isomorphism.
By theorem 4 and remark 7 we obtain
Corollary 2. If L is a graded Lie coalgebra in coLiegr there is a canonical quasi-isomorphism UL∨'−→
RHomCL(k, k).
We end this section by stating a fact from [44, section 1], which is in a sense Koszul dual to the
previous corollary. It can be proven similarly using theorem 13.
Theorem 14. If L is in coLiegr there is a canonical quasi-isomorphism CL '−→ RHomUL∨(k, k). This
results in a bigraded algebra isomorphism Hp(CL)q ∼= ExtpUL∨(k, k)q.
Chapter 3
Koszul Duality in Local
Commutative Algebra
As we discussed in the introduction, the behaviour of the homotopy Lie algebra π∗(φ) reflects the
character of the singularity defined by a local homomorphism φ : R → S. In quite a broad sense,
anything about the homotopy Lie algebra is for us something to do with Koszul Duality. In this chapter
we investigate a few ideas in this direction. However, many of directions remain to be properly explored,
and some important avenues will be completely ignored for now.
3.1 The Homotopy Lie algebra
In this section we explain how to assign a graded Lie algebra to a local ring R or a local homomorphism
φ : R→ S, using the background from section 2.8.
Recall that we are calling a dg algebra A a minimal model if it is a semi-free extension Q[X] of a
regular local ring Q with X in positive degrees, and A is absolutely minimal, so ∂(mA) ⊆ m2A. Whenever
A is a minimal model we define π∗(A) = Σ(mA/m2A) and π∗(A) = π∗(A)∨. Using the constructions in
section 2.7, minimal models come up for us in two ways:
1. In the absolute situation we take a minimal Cohen model Q[X]'−→ R, and then A = Q[X] is by
definition a minimal model. We define π∗(R) = π∗(A) and π∗(R) = π∗(A).
2. In the relative situation one builds a minimal Cohen model R→ R′[X]'−→ S for a homomorphism
φ : (R, k)→ (S, l). The fibre A = k⊗RR′[X] is then a minimal model, and we define π∗(φ) = π∗(A)
and π∗(φ) = π∗(A).
These definitions depend a-priori on the construction of A. Independence is dealt with by the main
theorem of this section, which we state now. It extends a classical theorem of Avramov [9, theorem 4.2],
which only deals with π≥2(R).
Theorem 15. If A is a minimal model with residue field k then π∗(A) is canonically a graded Lie
algebra over k, and there is a natural isomorphism Uπ∗(A) ∼= Ext∗A(k, k) of Hopf algebras with codivided
powers. In particular:
46
Chapter 3. Koszul Duality in Local Commutative Algebra 47
1. If (R, k) is a local ring then Uπ∗(R) ∼= Ext∗R(k, k).
2. If φ : (R, k) → (S, l) is a local homomorphism then Uπ∗(φ) ∼= Ext∗F (l, l) where F → l is any
augmented dg algebra model for k ⊗LR S → l.
The assertion about Hopf algebras with codivided powers will be explained later in this section. We
call π∗(R) and π∗(φ) the homotopy Lie algebra of R and of φ respectively. The theorem states that our
definition agrees with the classical one from, for example, [9].
Originally the homotopy Lie algebra was uncovered though Ext∗R(k, k). It was shown by Levin [89]
and Shoeller [109] that Ext∗R(k, k) is a Hopf algebra whose dual TorR∗ (k, k) is commutative with divided
powers in positive degrees. At the point the famous structure theorem of Milnor and Moore [99], and
Andre and Sjodin in positive characteristic [111], shows that Ext∗R(k, k) is the universal envelope of a
uniquely defined graded Lie algebra. Our approach is side step this structure theory and build π∗(R)
directly. From theorem 15 is follows a fortiori that Ext∗R(k, k) is a Hopf algebra with codivided powers.
We do not invoke the work of Levin or Schoeller.
To an absolutely minimal local dg algebra A we can associate a bigraded dg algebra gr(A) by setting
grwAi = (mwA/mw+1A )i. The upper grading is referred to as weight, the lower grading is the internal degree.
Using the fact that A is minimal there is an induced differential ∂ : (mwA/mw+1A )i → (mw+1
A /mw+2A )i−1.
It is convenient to fix a field k over which we will work implicitly from this point on. Let min-mod
be the full subcategory of minimal models with residue field k. And recall that the category q-semi-free
of quadratic semi-free dg algebras was defined in section 2.8. Since the associated graded algebra of a
regular local ring is a polynomial algebra, we obtain a functor
gr : min-mod −→ q-semi-free.
Through this, the fact that π∗(A) = (ΣmA/m2A)∨ = (Σgr1A)∨ is a graded Lie algebra is dealt with by
theorem 12.
The functor grA should be thought of roughly as taking the underlying graded Lie algebra of a
minimal Lie∞ algebra. That is, at this point we throw away some homotopical information which might
allow us to recover A from its homotopy Lie algebra.
Note that grA = C(π∗(A)). As in section 2.8 the functor we really care about is the completion
grA =∏w grwA. Then grA = C(π∗(A)) and following lemma combined with corollary 2 finishes the first
part of theorem 15.
Lemma 11. If A = Q[X] is a minimal model then there is a canonical isomorphism of graded algebras
Ext∗A(k, k) ∼= Ext∗grA(k, k).
Proof. The key point is that the hypotheses guarantee that if F → k is the minimal dg A module
resolution, then grF → k is the minimal dg grA module resolution. This can be seen from theorem
11 above, which says we may take F to be the acyclic closure Q[X]〈t, ΣX〉 of k, with a differential
that satisfies ∂(Σx) = x modulo (m2A + mA(t, ΣX)(1)), and ∂(t) = X0 where X0 is a minimal set of
generators for mQ. Then grA = kJX0, XK and by minimality grF = kJX0, XK〈t, ΣX〉 obtains a differential
∂ : mwA/mw+1A Fi → mw+1
A /mw+2A Fi−1 making it a graded dg module over grA. The differential of grF
Chapter 3. Koszul Duality in Local Commutative Algebra 48
still satisfies the conditions ∂(Σx) = x modulo (mgrA(t, ΣX)(1)) and ∂(t) = X0, and it’s well known that
this makes grF acyclic (see the proof of lemma 9).
We obtain a sequence of isomorphisms of graded vector spaces
Ext∗A(k, k) ∼= HomA(F, k) ∼= HomgrA(grF, k) ∼= Ext∗grA(k, k).
It remains to understand why this isomorphism Ext∗A(k, k) ∼= Ext∗grA(k, k) respects the algebra structure.
For this we have a commutative diagram
HomA(F, k) HomgrA(grF, k)
HomA(F, F ) HomgrA(grF, grF ).
∼=
φ 7→ grφ
' '
The map along the bottom of the diagram is not a chain map. However, it is easily seen to be an algebra
map which takes cycles to cycles. The product of Ext∗A(k, k) may be computed by lifting two elements
to cycles in HomA(F, F ) and composing them there, similarly for Ext∗grA(k, k). With this in mind, the
map φ 7→ grφ is good enough to witness the fact that the isomorphism above is one of algebras.
As is evident from the proof, the isomorphism Ext∗A(k, k) ∼= Ext∗grA(k, k) need not lift to quasi-
isomorphism between RHomA(k, k) and RHomgrA(k, k). In fact, we’ve seen already that RHomgrA(k, k)
is always formal, while RHomA(k, k) is usually not so.
Both 1 and 2 in theorem 15 are dealt with by
Lemma 12. Let A'−→ B be a quasi-isomorphism of local dg algebras with common residue field k. There
are canonical isomorphisms of graded algebras
Ext∗A(k, k) ∼= Ext∗B(k, k) ∼= Ext∗B
(k, k).
The proof of the lemma is standard. With this 1 is established using the quasi-isomorphism A =
Q[X]'−→ R. For 2, if F is a model for the derived fibre k ⊗L
R S then F is connected to A by a chain
of quasi-isomorphisms, where A = k ⊗R R′[X] is the minimal model arising from a minimal Cohen
model R → R′[X] → S for R → S. This is because R′[X] is a flat resolution of S over R. Therefore
Ext∗A(k, k) ∼= Ext∗F (k, k) by the lemma.
It remains to explain the assertion in theorem 15 about Hopf algebras with codivided powers.
TorA∗ (k, k) has the structure a graded coalgebra, dual to the Yoneda algebra structure on Ext∗A(k, k).
Taking an acyclic closure A → A〈U〉 '−→ k we see that TorA∗ (k, k) ∼= k〈U〉 by Gulliksen and Schoeller’s
minimality theorem 11. Thus TorA∗ (k, k) also has the structure of a (free) divided power algebra, with di-
vided powers supported on the ideal of positive degree elements. It is classical that this makes TorA∗ (k, k)
into a Hopf algebra with divided powers1, that is, the coproduct ∆ : TorA∗ (k, k) → TorA∗ (k, k)⊗2 is a
homomorphism of divided power algebras. We say that the dual Ext∗A(k, k) = TorA∗ (k, k)∨ is a Hopf
algebra with codivided powers.
1This generalises the work of Assmus, Levin and Schoeller [5, 87, 109] from the case that A is a local ring. The generalcases can be shown (for example) by using lemma 11 to reduce to the equicharacteristic case and then using lemma 13 to
present Uπ∗(A) as a model for Torgr(A)∗ (k, k) with this structure.
Chapter 3. Koszul Duality in Local Commutative Algebra 49
The isomorphism constructed in lemma 11 clearly preserves this structure. That is, it was explicitly
dual to an isomorphism TorA∗ (k, k) ∼= k〈U〉 ∼= Torgr(A)∗ (k, k) of divided power algebras.
The map from theorem 4 which was used to build the isomorphism in theorem 15 made use of the
twisted tensor product Cπ∗(A)⊗τUπ∗(A) ' k.
Lemma 13. Let L be a graded Lie coalgebra in coLiegr, and let τ : UL → CL be the acyclic twisting
cochain of theorem 13.
The natural divided paper structure is compatible with the twist by τ . That is, CL⊗τUL is a divided
power algebra with divided powers supported on the ideal(UL)
and thus CL → CL ⊗τ UL → k is an
acyclic closure.
Remark 8. The lemma works just as well for the bimodule resolution CL⊗τUL⊗τ CL.
Proof. It is a classical theorem the universal envelope of a graded Lie coalgebra is a Hopf algebra with
divided powers (this is a structural version of the Poincare-Birkhoff-Witt theorem, it is explicit in [32,
proposition 4.2]).
Let u be a primitive element of UL, with ∆(u) = 1⊗u+u⊗1. This implies ∆(u(n)) =∑i+j=n u
(i)⊗u(j), so
∂(u(n)) = (m⊗ 1)(1⊗ τ ⊗ 1)(1⊗∆)(1⊗ u)
=∑i+j=n
(m⊗ 1)(1⊗ τ ⊗ 1)(1⊗ u(i) ⊗ u(j))
= τ(u)⊗ u(n−1)
= ∂(u)u(n−1)
since τ(u(i)) = 0 unless i = 1. Since(UL)
is generated by primitives as a divided power ideal it follows
quite formally that ∂(u(n)) = ∂(u)u(n−1) for any u ∈(UL)
(this involves short a calculation for each of
the divided power axioms).
Note that through the augmentation there is an algebra homomorphism ε : UL → k → CL. Note
that τ is actually a derivation with respect to ε. It follows that the twisted differential satisfies
∂(uv) = (−1)|v′||u′′|τ(u′v′)u′′v′′
= (−1)|v′||u′′|(ε(u′)τ(v′)u′′v′′ + τ(u′)ε(v′)u′′v′′)
= (−1)|v′||u|τ(v′)uv′′ + τ(u′)u′′v
= u∂(v) + ∂(u)v
where we have used an informal Sweedler notation ∆(u) = u′ ⊗ u′′ and ∆(v) = v′ ⊗ v′′. (Actually τ is a
divided power derivation for the trivial divided power structure on the zero ideal of Cπ∗(A), so a similar
trick can be used to simplify the verification above.)
In all, this shows that CL⊗τUL is a dg divided power algebra. Since CL⊗τUL is acyclic and minimal
as a complex, the last assertion is clear.
It follows from the lemma that the natural divided power structure on Uπ∗(R) gives rise to the divided
power structure of TorCπ∗(A)∗ (k, k) ∼= k⊗Cπ∗(A) (Cπ∗(A)⊗τUπ∗(A)) ∼= Uπ∗(R). This isomorphism is dual
to the one given by theorem 4. We have finished the proof of theorem 15.
Chapter 3. Koszul Duality in Local Commutative Algebra 50
Example 4. Recall that Golod homomorphisms were discussed in section 2.7. Knowing from theorem
15 that our definition of the homotopy Lie algebra agrees with the usual one, it follows from [10, theorem
(3.4)] that φ : R→ S is Golod if and only if π∗(φ) is free as a graded Lie algebra.
Slightly extending the notion introduced by Avramov in [11] we also say that φ is generalised Golod
of level n if the graded Lie algebra π>n(φ) is free. Thus, the Golod condition coincides with generalised
Golod of level 0. It is not difficult to see that the conditions in proposition 5 are equivalent to being
generalised Golod of level 1 (see proposition 9 for more detail).
Remark 9. A local ring R is Koszul if its residue field has finite linearity defect, in the sense of [67].
Equivalently, the graded algebra gr(R) is Koszul in the classical sense of section 2.3 (see remark 1.10 of
op. cit.). It is also equivalent that π∗(R) is generated in degree 1 as a graded Lie algebra.
One can see using the proof of lemma 11 that R is Koszul if and only if gr(A) → gr(R) is a quasi-
isomorphism whenever A'−→ R is a minimal Cohen model. In particular, gr(R) can be recovered directly
from the homotopy Lie algebra π∗(R) by taking H0(Cπ∗(R)).
Deviations of a Local Ring or Homomorphism
Let R → R〈U〉 → k be an acyclic closure of the residue field of R. Classically, the deviations of R
have been defined as εn(R) = card(Un). By Gulliksen’s theorem 11 above there is an isomorphism
k〈U〉 ∼= TorR∗ (k, k). It follows that the deviations are uniquely determined by the well-known formula
PRk (t) =
∏i≥1(1 + t2i−1)ε2i−1(R)∏i≥1(1− t2i)ε2i(R)
,
see [12, remark 7.1.1]. By theorem 15 the same formula characterises the numbers dimk πi(R). Taking
a minimal model Q[X]'−→ R and choosing a minimal set X0 of generators for mQ, we recover another
theorem of Avramov that εn(R) = dimk πn(R) = card(Xn−1), see [9] or [12, proposition 7.2.3].
In [13, section 3] the deviations of a local homomorphism φ : (R, k)→ (S, l) are (essentially) defined
as εn(φ) = diml πn(φ). Theorem 15 allows one to express these invariants in terms of the Poincare series
of the dg algebra k ⊗LR S using exactly the same formula.
A great deal can be said about the behaviour of deviations, especially their growth. We will not
discuss this rich subject here. The reader may consult [65, 12, 13, 47, 63] for a beginning.
Functoriality of the Homotopy Lie Algebra
We end this section by describing briefly in what sense the homotopy Lie algebra is functorial. This is
essentially classical, but (except for the absolute homotopy Lie algebra) details are difficult to find in
the literature. So we sketch the arguments.
Suppose we have a commutative diagram of local homomorphisms
(R, k) (S, l)
(T, t) (U, u).
φ
α β
ψ
Following [16] we take a minimal Cohen factorisation T → T ′ → U , then we form the fibre product
T ′ ×U S. There is a natural homomorphism R → T ′ ×U S, and we can take a Cohen factorisation of
Chapter 3. Koszul Duality in Local Commutative Algebra 51
this to obtain R → R′ → T ′ ×U S. Finally, if the induced factorisation R → R′ → S is not minimal,
we can factor R′ = R′/(x) by a regular sequence x ⊆ mR′/m2R′
to obtain a minimal Cohen factorisation
R→ R′ → S.
We can then extend this construction to a diagram of minimal models
R′[X]
R R′[Σx,X] S
T T ′[Y ] U .
''
'
'
Having done this we set A = k ⊗R R′[X] and A = k ⊗R R′[X] and B = t⊗T T ′[Y ].
Theorem 16. The induced map
π∗(φ) = ΣmA/m2A ⊗l u
∼=←−− H∗(ΣmA/m2A
)⊗l u −→ ΣmB/m2B = π∗(ψ)
is a homomorphism of Lie coalgebras over u. It is independent of all choices made, and we denote it by
(α, β)∗ : π∗(φ)⊗l u→ π∗(ψ). By dualising we obtain a homomorphism (α, β)∗ : π∗(ψ)→ π∗(φ)⊗l u of
Lie algebras.
If we have another diagram of local homomorphisms
(T, t) (U, u)
(T ′, t′) (U ′, u′),
ψ
α′ β′
ψ′
then ((α′, β′)∗ ⊗u u′)(α, β)∗ = ((α′α), (β′β))∗ : π∗(φ)⊗l u′ → π∗(ψ′).
In many situations we can directly build compatible minimal Cohen models, see [101]. In this case
it is clear from theorem 12 that (α, β)∗ is a homomorphism of Lie coalgebras.
The theorem can be proven by invoking theorem 15 to work instead with divided power Hopf algebras,
after which this is essentially a classical theorem of Avramov [9, theorem 2.1].
We sketch some details now. Let R → R〈U〉 '−→ k and T → T 〈V 〉 '−→ t be acyclic closures. The
classical lifting property [64, lemma 1.8.6] allows us to lift k → t to R〈V 〉 → T 〈W 〉. By tensoring these
homomorphisms together we get a diagram of dg algebras
A R′[X]〈U〉 S〈V 〉
A R′[Σx,X]〈U〉 S〈V 〉
B T ′[Y ]〈W 〉 U〈W 〉.
' '
' '
' '
with quasi-isomorphisms as indicated. Now we use the fact that TorC∗ (c, u) is a functor from local dg
Chapter 3. Koszul Duality in Local Commutative Algebra 52
algebras (C, c) with a morphism c→ u to the category of Hopf algebras over u with divided powers, and
this functor takes quasi-isomorphisms to isomorphisms. The right column of the diagram is independent
of the choice of Cohen factorisations and maps between them, while the left column is independent of
the choice of acyclic closure and lift between them. Thus applying this functor establishes independence
and functoriality.
It remains to explain why the map described in the theorem is the same as the one induced by the
functor Tor(−)∗ (−, u) (after taking divided power indecomposables to recover π∗). This is a technical fact
which follows from the functoriality of the Tate model for the diagonal, proven in section 3.4. The result
we need is theorem 22 from that section, which is really a version of [21, Theorem 3.4].
A local homomorphism φ : (R, k) → (S, l) also gives rise functorially to a homomorphism of Lie
coalgebras which we will denote φ∗ : π∗(R)⊗k l→ π∗(S). It is dual to a homomorphism of Lie algebras
φ∗ : π∗(S)→ π∗(R)⊗k l.
3.2 The L.S. Category of a Ring Homomorphism
To a local homomorphism φ : R → S we will associate a sequence of numbers cati(φ) for i = 0, 1, 2...,
which, in brief, can be thought of as the Lusternik-Schnirelmann category of the i-connected cover of
the derived fibre of φ. The i-category can be a natural number, or it can be infinite. In this section we
prove that this invariant is well-defined and we establish some of its basic properties.
The ideas in this section are old. However, the definitions given below do not seem to appear elsewhere
in this generality, so the results are at least superficially new. Except for well-definedness of cat0 and
cat1, most of these these results follow without a great deal of fuss from known facts in the literature.
First, let us briefly describe the history of the L.S. category, beginning with topology. Much more
detail can be found in [46] or [68].
Let X be a topological space. A subspace U of X is called contractible in X if the inclusion U → X
is homotopic to a constant map. We then define the L.S. category of X to be the least integer m such
that X can be covered by m+ 1 open sets, each of which is contractible in X. This number is denoted
cat(X) = m. If no such m exists then X has infinite L.S. category. The normalisation is chosen so that
cat(X) = 0 precisely if X is contractible. Note also that cat(Sn) = 1 for the n-dimensional sphere Sn.
This invariant was introduced by Lusternik and Schnirelmann in [93], where it is shown that if X
is a closed, smooth manifold then any smooth function on X has strictly more than cat(X) critical
points. The L.S. category was imported to rational homotopy theory in the thesis of Lemaire [86].
After the ground-breaking work of Felix and Halperin in [43], the central position of L.S. category in
rational homotopy theory was cemented. In short, a simply connected space X is called rational if all of
the homotopy groups πi(X) are rational vector spaces, and crucially rational spaces can be represented
algebraically by their Sullivan models, which are certain semi-free dg commutative algebras over Q. Felix
and Halperin showed that for rational spaces L.S. category can be computed algebraically in terms of
their Sullivan models. The definition given below for ring homomorphisms is a direct transcription of
their characterisation. From this they were able to deduce the Mapping Theorem: if X → Y is a map
of simply connected rational spaces such that each πi(X) → πi(Y ) is injective, then cat(X) ≤ cat(Y ).
Further, they used the Mapping Theorem to prove the beautiful Dichotomy Theorem: if X is a simply
Chapter 3. Koszul Duality in Local Commutative Algebra 53
connected rational space of finite category then the sequence of numbers dimQ πi(X) is either eventually
zero, or it must grow exponentially.
Around the time of [43] Avramov and Roos began to uncover deep connections between local commu-
tative algebra and rational homotopy theory. This has led to a great deal of collaboration and progress
in both areas, some of which is described in the surveys [9] and [6]. As an example, Avramov in [13] has
since established the local commutative algebra analogue of the Dichotomy Theorem: if φ : R→ S is a
local homomorphism of finite category, then either φ is complete intersection or the deviations εi(φ) grow
exponentially (see also [21, corollary 5.5]). Another highlight is the so-called Five Author Paper of Felix,
Halperin, Jacobsson, Lofwall and Thomas [44], in which the authors use L.S. category to constrain the
structure of the homotopy Lie algebra in both rational homotopy theory and local commutative algebra
(see theorem 18 below).
The Definition of L.S. Category
We now define the i-category of a local homomorphism φ : R → S in stages, starting with category for
dg algebras.
If (A,mA, k) is a local dg algebra then the category of A is the least integer m such that the quotient
A → A/mm+1A factors as a split monomorphism followed by a surjective quasi-isomorphism. In other
words, cat(A) is the least m for which a diagram of local dg algebras exists
A B
A/mm+1A
ι
'
ρ
in which the triangle commutes and ρι = 1. If no such m exists then by definition A has infinite category.
While the definition makes sense for any local dg algebra, we will usually only apply it when A is
a minimal model. That is, when ∂(mA) ⊆ m2A and A has the form Q[X] for a regular local ring Q
and a graded set X in strictly positive degrees. It is these dg algebras for which category has the best
properties. However, the extra generality is necessary and we will make no such assumption below unless
explicitly mentioned.
Remark 10. Note that if cat(A) ≤ m then the maximal ideal mH∗(A) of H∗(A) satisfies mm+1H∗(A) = 0,
and in particular the product of more than m positive degree elements must vanish. It is useful to have
a name for this nilpotency: if B is a local graded algebra then the Loewey length of B is the (possibly
infinite) number
``(B) = infi ≥ 0 : miB = 0.
So, the observation above may be rewritten as cat(A) ≥ ``(H∗(A))− 1. However, finiteness of category
is a much stronger and more structural nilpotence condition than this. It says that, up to homotopy, A
is a retract of A/mm+1A .
Remark 11. There is by proposition 4 a minimal semi-free model for the quotient A → A[Y ]'−→
A/mm+1A , and this model is unique up to isomorphism. If cat(A) = m then in fact this extension
A→ A[Y ] admits a retract. This is because by the lifting property of lemma 5 there is an A algebra lift
Chapter 3. Koszul Duality in Local Commutative Algebra 54
A[Y ] → B, which makes A → A[Y ] split via B. Hence, to establish the inequality cat(A) ≤ m we will
always take the minimal model A→ A[Y ]'−→ A/mm+1
A and attempt to construct a splitting A← A[Y ].
Remark 12. Let m ≤ n be two natural numbers and suppose that a retract exists as in the above
diagram, demonstrating that cat(A) ≤ m. Then if A → A[Y ]'−→ A/mn+1
A is a minimal model there is
also a retract A← A[Y ]. This follows readily from the existence of the A algebra lift in the diagram
A[Y ] B
A/mn+1A A/mm+1
A .
' '
Let A be a minimal model. We define the i-connected cover of A to be the dg algebra
A(i) = A/([mA]<i)
where ([mA]<i) is the ideal in A generated by all elements of mA with degree less than i.
The name i-connected cover comes by analogy with topology. The homomorphism A→ A(i) induces
a map π∗(A(i)) → π∗(A) of graded Lie algebras, and by construction this will give an isomorphism
π∗(A(i)) ∼= π>i(A). In other words, the effect of factoring by ([mA]≤i) is to remove the homotopy groups
π<i(A).
Now suppose that (R,m, k) and (S, n, l) are local rings, and let φ : R→ S be a local homomorphism.
Let R → R′[X] → S be a minimal Cohen model of φ, as in section 2.7. In particular, the fibre
Q = R′/mR′ is regular and A = Q[X] = k ⊗R R′[X] is a model for the derived fibre k ⊗LR S. For a
natural number i we define the i-category of φ to be
cati(φ) = cat(A(i))
In other words, cat0(φ) is the category of Q[X], while cati(φ) is the category of l[X≥i] = l[X]/(X<i) for
any i > 0.
We also define the absolute i-category of a local ring by taking a minimal modelA = Q[X]'−→ R. Then
exactly as before cati(R) = cat(A(i)) by definition. Thus, the inequality cati(R) ≤ m says intuitively
that the i-connected cover of R satisfies the strong (m+ 1)-nilpotence condition described above.
If R is equicharacteristic we recover the absolute situation by considering a coefficient field k → R.
That is, we have cati(R) = cati(k → R) for all i. Conversely, if φ : R→ S is a flat local homomorphism
with fibre F = k ⊗R S then it is easy to see that cati(φ) = cati(R) for all i.
Properties of L.S. Category and Examples
Recall that when the extension k → l is separable any two minimal Cohen factorisations of R → S are
isomorphic, and hence A is determined up to isomorphism. In the general case, we need to establish
that the numbers cati(R) and cati(φ) are well-defined.
Theorem 17. For all i the numbers cati(R) and cati(φ) are independent of the choice of minimal models.
If R is a quotient of a regular local ring Q then the i-category may be computed using a minimal model
Q[X]'−→ R (without completing R). Similarly, if φ admits a minimal regular factorisation R→ T → S
then the i-category may be computed using a minimal model T [Y ]'−→ S (without completing S).
Chapter 3. Koszul Duality in Local Commutative Algebra 55
Note that the last statement applies in particular to surjective homomorphisms. To be more precise,
the assertion is that cati(φ) = cat(T [Y ](i)) for all i, where T is the fibre T/mT , even though cati(φ) is
computed using a Cohen factorisation of the semi-completion φ : R→ S.
Let us point out that well-definedness of cati for i ≥ 1 is essentially proven by Avramov and Iyengar
in [21, proposition 2.4]. The main difficulty will be in dealing with cat0 and cat1. Since the proof of the
theorem is long we delay it until the end of this section. Instead we start by discussing some properties
of this invariant.
Firstly, the following fact is an immediate consequence of the mapping theorem, which will be stated
later in this section (and proven in section 3.3).
Proposition 6. For any local homomorphism φ : R→ S, the i-category decreases as i increases
cat0(φ) ≥ cat1(φ) ≥ cat2(φ) ≥ cat3(φ) · · ·
Remark 13. In [21] Avramov and Iyengar consider the weak category of a local homomorphism φ :
R→ S, which they define as follows
wcat(φ) = infm ∈ N : ``(H∗(A
(i)) ≤ m+ 1 for all i ≥ 2,
where A = k⊗RR′[X] is the dg algebra constructed as above from a minimal Cohen model R→ R′[X]→S. By the previous proposition wcat(φ) ≤ cat2(φ). Avramov and Iyengar show that finiteness of weak
category suffices to establish good properties of the homotopy Lie algebra π∗(φ), and they prove that
almost small homorphisms have finite weak category. In fact, it will be a consequence of the mapping
theorem that almost small homomorphisms have finite 1-category. See corollary 3 for details and the
definition of almost small homomorphisms.
Unsurprisingly, category and Loewy length are closely connected invariants. We have already ob-
served that cat(A) ≥ ``(H∗(A))− 1. In the other direction, the following lemma is extremely useful for
bounding the i-category of a homomorphism.
If B is a local dg algebra then ``(B) is by definition the Loewy length of the underlying graded
algebra. This is certainly not a quasi-ismorphism invariant. Rather, it is a property of a particular
model for B.
Lemma 14. Let A and B be local dg algebras.
1. If there is a surjective quasi-isomorphism A→ B then cat(A) ≤ ``(B)− 1.
2. If A is a retract of B then cat(A) ≤ cat(B).
Proof. For 1 assume that ``(B) = m + 1 is finite. Let A → A[Y ]'−→ A/mm+1
A be a minimal model.
Because mm+1B = 0 the homomorphism A→ B factors through A/mm+1
A , as in the diagram
A A[Y ]
A/mm+1A B.
'
Since A→ B is a surjective quasi-isomorphism the standard lifting property from lemma 5 produces an
A algebra splitting as showing in the diagram. A similar lifting argument easily establishes 2.
Chapter 3. Koszul Duality in Local Commutative Algebra 56
Note that cat0(φ) = cat1(φ) for surjective homomorphisms. In general, cat0(φ) is will be infinite
unless the fibre S = S/mS is Artinian, because ``(S) ≤ cat0(φ) + 1.
Example 5. Consider the homomorphism f : kJf1, ..., fcK → kJx1, ..., xnK where f1, ..., fc is a regular se-
quence in (x1, ..., xn)2. Then f is flat and the fibre A = kJx1, ..., xnK/(f1, ..., fc) is a complete intersection
of codimension c. One can easily build a minimal Cohen factorisation
kJf1, ..., fcK→ kJf1, ..., fc, y1, ..., ynKyi 7→xi−−−−−→ kJx1, ..., xnK.
Since the kernel of kJf1, ..., fc, y1, ..., ynK → kJx1, ..., xnK is generated by the regular sequence (f1(y) −f1), ..., (fc(y) − fc) the homomorphism f is complete intersection in the sense of [13]. For a minimal
model we have the Koszul complex kJf1, ..., fc, y1, ..., ynK[Z]'−→ kJx1, ..., xnK where Z = z1, ..., zc is a
set of degree one variables with ∂(zi) = fi(y)− fi. Hence
cat0(f) = cat(kJy1, ..., ynK[Z]), cat1(f) = cat(k[Z]) and cat≥2(f) = cat(k).
Thus cat≥2(f) = 0, and since k[Z] is an exterior algebra on c generators cat1(f) = c is the codimension
of A. I claim that cat0(f) = ``(A) − 1. Suppose that m + 1 = ``(A) is finite (this is equivalent to
n = c). Then applying lemma 14 to the surjective quasi-isomorphism kJy1, ..., ynK[Z] → A shows that
cat0(f) ≤ ``(A)−1. As we have already mentioned the reverse inequality always holds, so this concludes
the calculation of cati(f) for all i. Note that if the fi come from homogeneous forms in the standard
graded polynomial algebra k[x1, ..., xn] then (assuming n = c) the number cat0(f) = ``(A) − 1 is the
socle degree of A, which is equal to the product deg(f1) · · · deg(fc).
On the other hand cat1(φ) is often finite even when the fibre S is not Artinian. As a consequence of
the next proposition, this is always the case if R is regular.
Proposition 7. For any minimal model A such that A0 is a field we have cat(A) ≤ sup(A).
For a local ring R we have cat0(R) = ``(R)− 1 and cat1(R) ≤ codepth(R). In particular cati(R) is
always finite for i > 0.
For a local homomorphism φ : R→ S we have cat1(φ) ≤ fdRS + edimS/mS, where fdRS is the flat
dimension of S as an R module.
In the statement we have used the notation codepth(R) = edim(R)− depth(R).
The analogous statement in rational homotopy theory is that a simply connected space X with
finite dimensional cohomology H∗(X;Q) has finite rational category. This local algebra version is also
essentially well known, compare [13, lemma 3.8] or [21, theorem 5.7].
Proof. Recall that m = sup(A) is by definition supi ∈ N : Hi(A) 6= 0. Define a graded subspace
I ⊆ A by setting Ii = Ai when i > m and Im = ∂(Am+1) and Ii = 0 when i < m. By construction I is
an acyclic ideal of A and hence A → A/I is a surjective quasi-isomorphism. But mm+1A/I = (A/I)m+1
>0 ⊆(A/I)≥m+1 = 0 so by lemma 14 cat(A) ≤ ``(A/I) ≤ m.
For the second assertion we take a minimal model Q[X] → R. Then by lemma 14 cat0(R) =
cat(Q[X]) ≤ ``(R)− 1, and the other inequality is clear. Using the previous part cat1(R) = cat(k[X]) ≤sup(k[X]) = pdQ(R) = codepth(R).
For the last assertion, take a minimal Cohen model R → R′[X]'−→ S and write Q = k ⊗R R′. By
definition of minimality dimQ = edimS/mS. Since Q[X] is a model for k⊗LRS we have supQ[X] = fdRS.
Chapter 3. Koszul Duality in Local Commutative Algebra 57
Now the 1-category of φ is cat(l[X]), which is bounded above by sup(l[X]). If K is the Koszul complex
over Q on a minimal set of generators for mQ, then there is a quasi-isomorphism of complexes K ⊗QQ[X]
'−→ l[X]. A simple spectral sequence argument now shows that sup(k[X]) = sup(K ⊗Q Q[X]) ≤fdRS + dimQ = fdRS + edimS/mS. This establishes the inequality cat1(φ) ≤ fdRS + edimS/mS.
Remark 14. The bound cat1(φ) ≤ fdRS + edimS/mS can be improved dramatically, because the
estimate of sup(l[X]) in the proof above was very crude. The actual computation of sup(l[X]) is doable,
but since the argument is long we leave it for now and content ourselves with the bound given in the
proposition.
The extent to which cat(A) only depends only on the Lie algebra π∗(A) is an interesting question.
Example 5 shows that in general cat(A) cannot be computed from π∗(A): in the notation there π∗(f)
only depends on the quadratic part of f while cat0(f) does not. However a similar argument to the
previous proposition yields the following result from the Five Author Paper [44].
Proposition 8. Let A be a complete minimal model and let L = π∗(A) be its homotopy Lie algebra.
Then cat(A) ≤ gldim(UL) where gldim(UL) is the global dimension supi : ExtiUL(k, k)∗ 6= 0
.
Proof. We sketch the argument in [44]. If Ext>gUL(k, k)∗ = 0 then by theorem 14 also H>g(grA)∗ = 0. Let
V ⊂ mgA be any Q submodule whose image in mgA/mg+1A is a complement to the kernel of ∂ : grg(A) →
grg+1(A). Then I = V +mg+1A is an ideal in A and gr(I) is acyclic by construction. By lemma 1 I itself
is acyclic, and just as in the proof above this forces cat(A) ≤ g.
The authors of [44] also show that depth(UL) ≤ cat(A), and that if this is an equality then so is
cat(A) = gldim(UL).
An important theme in what follows is that it is desirable to have a homomorphism of finite i-
category, and thus it is desirable to have theorems which bound category from above. This theme came
up in the historical discussion above. For example, in the Five Author Paper it is shown that there are
interesting consequences for the structure of the homotopy Lie algebra of a minimal model with finite
category.
Theorem 18 (Felix-Halperin-Jacobsson-Lofwall-Thomas [44, theorems B and C]). If A = k[X] is a
minimal semi-free dg algebra over a field k such that cat(A) is finite, then the homotopy Lie algebra
π∗(A) has finite dimensional radical, and
dim(rad(π∗(A))even) ≤ cat(A).
In particular this bounds the dimension of the centre of π∗(A). Let R be a local ring. By results of
Avramov and Sun [23, theorem 5.3] and Avramov, Gasharov and Peeva [17, theorem 5.3] the complexity
cxR(M) of any module M with finite complete intersection dimension is bounded above by the dimension
of the degree 2 part of the centre of π∗(R). It follows in particular that cxR(M) ≤ cat0(R). Hence results
which bound category above are potentially useful. A relative version of this fact involving π∗(φ) is also
conceivable.
Another reason why it is desirable to have a homomorphism of finite category is that it allows one to
construct long exact sequences of homotopy Lie algebras (see theorem 25 and the exact sequence (3.2)
below it). This is proven in unpublished work of Avramov, see the discussion in [9] and [6]. Details for
Chapter 3. Koszul Duality in Local Commutative Algebra 58
the case of a flat homomorphism are in [7]. Some details are also given below in section 3.5, where we
discuss how to obtain long exact sequences in more general situations.
Aside from the question of finiteness, the values taken by i-category are interesting in their own right.
The following proposition summarises in a few cases what it means for a homomorphism to have small
i-category.
Recall that complete intersection homomorphisms, quasi-complete intersection (qci) homomorphisms
and Golod homomorphisms were defined in section 2.7. The generalised Golod conditions were defined
in example 4. In short, φ is generalised Golod of level i if the graded Lie algebra π>i(φ) is free.
Proposition 9. Let φ : (R,m, k)→ (S, n, l) be a local homomorphism, then
1. cat0(φ) = 0 if and only if φ is flat and its fibre is a field.
2. cat1(φ) = 0 if and only if φ is weakly regular.
3. cat2(φ) = 0 if and only if φ is complete intersection.
4. cat3(φ) = 0 if and only if φ is quasi-complete intersection and either φ is a complete intersection
or k has characteristic zero.
5. cat0(φ) ≤ 1 if and only if φ is Golod.
6. More generally cati(φ) ≤ 1 if and only if φ is generalised Golod of level i.
Note that if the extension k → l is separable 1 means that φ is formally etale, and if k → l is an
isomorphism this means the completion φ : R → S is an isomorphism. Similarly, in the separable case
2 means that φ is formally smooth.
Proof. The equivalences 1, 2 and 3 are all easy consequences of the definitions (but see [21, corollary
5.5] for details relevant 3).
For 4, assume that char(k) = 0 and that φ is a quasi-complete intersection. Then we have a minimal
Tate model R→ R′〈X〉 '−→ S with X = X≤2. Since char(k) = 0 this is also a minimal Cohen model, so
cat3(φ) = 0.
Conversely, suppose that cat3(φ) = 0 and φ is not complete intersection. Then we have a minimal
Cohen model R → R′[X]'−→ S with X = X≤2. The key point is that the hypothesis guarantees that
R′[X] is minimal as a complex of R′ modules. Suppose towards a contradiction that k has characteristic
p > 0. Since φ is not complete intersection X2 contains some element x. If R′ is equicharacteristic
then we easily arrive at the contradiction that [xp] is a nontrivial class in H2p(R′[X]), even though
H∗(R′[X]) ∼= S. If R′ is not equicharacteristic, use Cohen’s theorem to obtain a surjection Q→ R′ from
a regular local ring Q. Build a minimal model Q → Q[Y ]'−→ R′. By uniqueness of minimal models,
this extends to a minimal model Q[Y ] → Q[Y,X]'−→ S which is a lift of R′[X]. Now xp is a cycle in
(Q/p)[Y,X]2p, and it cannot be a boundary since it maps to xp in (R′/p)[X]2p, which is nontrivial since
(R′/p)[X]2p is minimal as a complex. This is a contradiction since (Q/p)[Y,X] ' (Q/p)⊗LQ S can have
homology only in degrees 1 and 2. It follows that char(k) = 0, and the Tate model R′[X] = R′〈X〉 shows
that φ is a quasi-complete intersection.
It remains to establish the equivalences 5 and 6. These will be straightforward corollaries of the
results in [10]. We treat 6 since 5 is a special case.
Chapter 3. Koszul Duality in Local Commutative Algebra 59
Let R → R′[X]'−→ S be a minimal model and define A = k ⊗R R′[X](i), so that cati(φ) = cat(A).
According to [10, theorem 3.4] A is a Golod dg algebra if and only π∗(A) ∼= π>i(φ) is free, if and only
if H∗(mA) → H∗(mA/m2A) = mA/m
2A = π∗(A) is injective (this map may be thought of as dual to the
Hurwicz of A). We will show that this last condition is equivalent to cat(A) ≤ 1. By definition if
cat(A) ≤ 1 then H∗(mA) is a retract of H∗(mA/m2A) = mA/m
2A, so this direction is clear. If H∗(mA) →
mA/m2A is injective, chose any l vector space compliment Z in mA/m
2A to the image of H∗(mA). The
composition A → A/m2A → A/(m2
A, Z) is then a surjective quasi-isomorphism, so cat(φ) ≤ 1 by lemma
14.
The proposition should give some intuition on the behaviour of category. It may be instructive to
display the conditions cati(φ) = j in an array:
0 1 · · ·cat0 flat with field fibre Golod
cat1 weakly regular gen.Golod of level 1
cat2 complete intersection gen. Golod of level 2
cat3 qci with char(k) = 0 gen. Golod of level 3...
. . .
As we move downwards we see increasingly relaxed regularity conditions. However, even going down the
first column we immediately run into the conjectures stated by Quillen in [106]. The first of these conjec-
tures has been settled by Avramov [13], and using these ideas Avramov and Iyengar prove [21, corollary
5.5], showing that the first column of the table stabilises almost immediately under the assumption
wcat(φ) <∞: the conditions cati(φ) = 0 all imply that φ is complete intersection.
Moving to the right in the table one obtains increasingly relaxed Golod-like conditions. Intuition
from Koszul duality is useful here: a trivial commutative algebra is Koszul dual to a free Lie algebra. As
the derived fibre of φ becomes less trivial its category increases, and the Koszul dual Lie algebra π∗(φ)
becomes less free.
We leave the absolute version of the proposition to the reader. In short, cat0(R) = 0 only when R is
a field; cat1(R) = 0 means that R is regular; and cat2(R) = 0 means that R is a complete intersection.
The Golod terminology is slightly miss-matched: the strong condition cat0(R) ≤ 1 is satisfied only when
the multiplication of R is trivial, that is, when m2R = 0. Instead, R is Golod if and only if cat1(R) ≤ 1.
Using a result of Wiebe [121] one can also characterise complete intersections using cat1.
Proposition 10. A local ring (R,m, k) is a complete intersection if and only if cat1(R) = codepth(R).
Since in general 0 ≤ cat1(R) ≤ codepth(R), the proposition puts regular rings and complete inter-
sections at opposite ends of a scale with the different values of cat1(R) interpolating. On this scale the
Golod condition is as close to regular as possible. The proposition was suggested as a possibility by
Luchezar Avramov.
Proof. Let c = codepth(R). If R is a complete intersection the computation of cat1(R) is as in example 5.
Let Q→ R be a minimal model. The kernel is generated by a regular sequence f = f1, ..., fc in Q. Thus
there is a minimal model Q[X]'−→ R with X = X1 and ∂(X1) = f . From here cat1(R) = cat(k[X]) = c
since k[X] is an exterior algebra on c generators.
Chapter 3. Koszul Duality in Local Commutative Algebra 60
For the converse it is convenient to use the following version of Wiebe’s criterion, taken from [35,
theorem 2.3.15]. The ring R is a complete intersection if and only if H>0(K)c 6= 0, where K is the Koszul
complex over R on a minimal generating set for m.
Assume R is not a complete intersection and let Q[X]'−→ R be a minimal model. Then cat1(R) =
cat(k[X]). Note that sup(k[X]) = pdQ(R) = c. If we let I be the ideal in k[X] with Ii = k[X]i for
i > c and Ic = ∂(k[X]c+1), then k[X] → k[X]/I will be a surjective quasi-isomorphism. I claim that
``(k[X]/I)− 1 < c, using part 1 of lemma 14 this will show that cat(k[X]) < c and complete the proof
of the proposition.
Let z1, ..., zc be in mk[X]/I . To establish the claim we must show that the product z1...zc is zero. For
degree reasons it is clear that z1...zc = 0 if any of the zi are in k[X]≥2. So we assume all the zi have
degree 1. In particular they are automatically cycles.
There is a chain of quasi-isomorphisms k[X] ' K, so H∗(k[X]) ∼= H∗(K). Since R is not a complete
intersection the class [z1]...[zc] vanishes in H∗(k[X]) by Wiebe’s criterion. Then z1...zc is a boundary in
k[X], and by definition this means z1...zc is in Ic. Therefore z1...zc = 0 in k[X]/I, and we are done.
Example 6. The generic short local ring is generalised Golod of level 2. Building on work of Conca
[40], Avramov, Iyengar and Sega [22] show that for a generic local ring (R,m, k) satisfying m3 = 0 there
is a quadratic hypersurface ring (Q, n, k) and a surjective Golod homomorphism φ : Q→ R whose kernel
is contained in n2. Since φ is Golod it is small [8, theorem 3.5], and it follows that there is a short exact
sequence of homotopy Lie algebras
0→ π∗(φ)→ π∗(R)→ π∗(Q)→ 0
(this will be explained in section 3.5). Note that π∗(Q) is edim(Q) = edim(R) dimensional in degree
1 and one dimensional in degree 2, after which it vanishes. Now the fact that a sub-Lie algebra of
a free Lie algebra is free implies π>2(R) is free (see [86, proposition A.1.10] and remark 16 below).
Generically m2 6= 0 so φ is not an isomorphism, and it follows that π>1(R) is not free. At this point we
can completely calculate the sequence cati(R):
cat0(R) = 2, cat1(R) = 2 and cati(R) = 1 for all i ≥ 2.
On the other hand by results of Anick [3] and Avramov [11] short local rings exist which are not
generalised Golod. These rings satisfy cati(R) = 2 for all i.
At the beginning of this section we alluded to the fact that theorems in rational homotopy theory have
often produced analogous theorems in local commutative algebra, and vice versa. The local commutative
algebra version of the Mapping Theorem has apparently been missing for some time. We will give a
proof in this thesis.
Theorem (theorem 20 below). If φ : A→ B is a surjective homomorphism of minimal models then
catA ≥ catB.
Keeping in mind the inequalities sup(k[X]) ≥ cat(k[X]) ≥ ``H∗(k[X])+1, this generalises a theorem
of Avramov and Iyengar, which may be stated as follows.
Chapter 3. Koszul Duality in Local Commutative Algebra 61
Theorem 19 (Avramov-Iyengar [19, theorem 1.2]). If k[X]→ k[Y ] is a surjection of minimal semi-free
dg algebras over a field k, then
sup k[X] ≥ ``H∗(k[Y ]) + 1.
Proof of Well-Definedness
To finish this section we will return to the proof of theorem 17, which says that the numbers cati(φ)
associated to a local homomorphism φ : (R,m, k)→ (S, n, l) are well-defined.
The main point is the following lemma, which we will set up now. If Q → P is a surjection of
regular local rings, then the kernel is automatically generated by a regular sequence x in mQ rm2Q. Let
Q → Q[t] → P be the Koszul complex which resolves P over Q, where t is a set of variables in degree
1 with ∂(t) = x. Given a minimal semi-free extension Q[t] → Q[t,X], we can apply − ⊗Q[t] P to get a
(quasi-isomorphic) minimal semi-free extension P → P [X].
Lemma 15. In the situation just described, catQ[t,X] = catP [X].
To establish this we first separate out the key technical observation.
Lemma 16. The natural surjection Q[t,X]/mm+1Q[t,X] → P [X]/mm+1
P [X] is a quasi-isomorphism
Proof. We can simplify the situation by noting that it suffices to treat the codimension one case.
In other words, we may assume that P = Q/(x) for a single non-zero-divisor x in mQ r m2Q. In
particular the Koszul complex Q[t] looks like tQx∂t−−→ Q. As a Q module Q[t,X]/mm+1
Q[t,X] is the di-
rect sum t(Q[X]/mmQ[X]) + (Q[X]/mm+1Q[X]), but the differential has components crossing in both direc-
tions. To fix this we filter by setting F(i)(Q[X]/mm+1Q[X]) = t(Q[X]/mmQ[X])≤i+1 + (Q[X]/mm+1
Q[X])≤i. We
also filter P [X]/mm+1P [X] by degree, so F(i) = (P [X]/mm+1
P [X])≤i. By lemma 1 it suffices to show that
gr(π) : gr(Q[t,X]/mm+1Q[t,X]) → gr(P [X]/mm+1
P [X]) is a quasi-isomorphism. But gr(Q[t,X]/mm+1Q[t,X]) looks
like
t(Q[X]/mmQ[X])x∂t−−−→ (Q[X]/mm+1
Q[X])
with only the differential x∂t as indicated, and gr(P [X]/mm+1P [X]) has no differential at all. It is clear
that the cokernel of x∂t is P [X]/mm+1P [X]. Thus to show that gr(γ) is a quasi-isomorphism we must show
that multiplication Q[X]/mmQ[X]
x−→ Q[X]/mm+1Q[X] is injective. This map diagonalises: both sides split
into a direct sum of Q modules generated by monomials in X, and multiplication by x respects this
decomposition. So for a monomial M of weight i < m in SymiQ(X) ∼= (X)i/(X)i+1, we must show that
MQ/MmjQx−→ MQ/Mmj+1
Q is injective, where i+ j = m. But this is isomorphic to Q/mjQx−→ Q/mj+1
Q .
Finally, this is injective because x is a non-zerodivisor in the associated graded ring grmQ, which is a
polynomial algebra since Q is regular.
Proof of lemma 15. First we deal with the inequality catQ[t,X] ≤ catP [X]. This is formal and applies
to any sequence of extensions Q→ Q[t]→ Q[t,X] with a quasi-isomorphism Q[t]→ P , with no need for
regularity. Suppose that catP [X] ≤ m is finite, and construct minimal models Q[t,X]→ Q[t,X, Y ]'−→
Q[t,X]/mm+1Q[t,X] and P [X] → P [X,Z]
'−→ P [X]/mm+1P [X]. By the lifting property of semi-free extensions
Chapter 3. Koszul Duality in Local Commutative Algebra 62
there is a Q[t,X] algebra map α : Q[t,X, Y ]→ P [X,Z] as in the following commutative diagram
Q[t,X, Y ] P [X,Z]
P [X]/mm+1P [X].
α
'
By definition of category the map P [X] → P [X,Z] is split by some β : P [X,Z] → P [X]. Then we use
the lifting property again on the following diagram of Q[t,X] algebra homomorphisms
Q[t,X] Q[t,X, Y ]
P [X],
'βα
to produce a splitting Q[t,X, Y ]→ Q[t,X]. This means that cat(Q[t,X]) ≤ m.
The inequality catQ[t,X] ≥ catP [X] relies more on the specific situation. So suppose catQ[t,X] ≤ mis finite. First observe that using lemma 16 and the commutative square
Q[t,X, Y ] Q[t,X]/mm+1Q[t,X]
P [X,Y ] P [X]/mm+1P [X],
'
' '
the homomorphism P [X,Y ]→ P [X]/mm+1P [X] is a quasi-isomorphism. Now take a minimal modelQ[t,X]→
Q[t,X, Y ]'−→ Q[t,X]/mm+1
Q[t,X] with a splitting Q[t,X, Y ]→ Q[t,X]. We apply −⊗Q[t] P to get the dia-
gram
P [X] P [X,Y ]
P [X]/mm+1P [X],
'
and this is witness to the inequality catP [X] ≤ m.
Proof of theorem 17. We treat only the relative situation, for a local homomorphism φ : R → S. The
absolute statement is similar and slightly simpler.
Assume that we have two minimal Cohen factorisations connected by a deformation as in the diagram
R′
R A S
R′′.
The fibres R′ = R′/mR′, R′′ = R′′/mR′′, and A = A/mA are all regular local rings.
Recall from the discussion in section 2.7 that up to isomorphism we may assume R′ = A/(x) and
Chapter 3. Koszul Duality in Local Commutative Algebra 63
R′′ = A/(y), where x and y are regular sequences in A which both descend to bases for (I + mA +
m2A)/(mA + m2
A), with I being the kernel of A → S (see also [16, (1.5)]). Note that x and y can be
extended in the same way to a minimal generating set for I, so we may build a minimal model for
A → S of the form A → A[s,X] = A[t,X]'−→ S, where s and t are sequences of variables in degree 1
with ∂(s) = x and ∂(t) = y, and where s ∪ X and t ∪ X both freely generate A[s,X] = A[t,X]. We
obtain minimal models
R′[X] = A[s,X]/(x, s)'←−− A[s,X] = A[t,X]
'−−→ A[t,X]/(y, t) = R′′[X].
For i ≥ 2 we have identifications R′[X]/(mR′ , X<i) = A[s,X]/(mA, s,X<i) = A[t,X]/(mA, t,X<i) =
R′′[X]/(mR′′ , X<i). Hence in this case the dg algebras used to compute cati(φ) are all isomorphic. This
argument appears already in [21, proposition 2.4].
For i = 0 we pass to fibres
R′[X]'←−− A[s,X] = A[t,X]
'−−→ R′′[X].
Then the lemma says that cat(R′[X]) = cat(A[s,X]) = cat(A[t,X]) = cat(R′′[X]). This shows that
cat0(φ) is well-defined.
To deal with i = 1, let z be a regular sequence in A whose image in mA/m2A
is a basis for a complement
to the inclusion (I+mA+m2A)/(mA+m2
A) → mA/m2A
. By construction (z)R′ = mR′ and (z)R′′ = mR′′ .
We obtain quasi-isomorphisms
l[s,X]/(s)'←−− (A/z)[s,X] = (A/z)[t,X]
'−−→ l[t,X]/(t).
At this point our notation could become misleading: l[X] = l[s,X]/(s) and l[X] = l[t,X]/(t) are
isomorphic as algebras but their differentials will be different in general. In any case, we can use the
lemma again to conclude that cat(l[s,X]/(s)) = cat((A/z)[s,X]) = cat((A/z)[t,X]) = cat(l[t,X]/(t)).
This establishes well-definedness of i-category for all i.
Knowing this, the final part of the theorem is dealt with by the following proposition.
Proposition 11. If A is a minimal model then cat(A) = cat(A).
Proof. Say A = Q[X] for some regular local ring Q, so that A = Q[X].
Note that if Q is a field then A = A and the proposition is obvious. This is the case when A is being
used to compute cati(φ) or cati(R) for any i > 0. The difficultly is in dealing with i = 0.
If cat(A) ≤ m then one can build a minimal model A→ A[Y ]'−→ A/mm+1
A with a splitting A[Y ]→ A.
Simply applying −⊗Q Q shows that cat(A) ≤ m. Hence cat(A) ≥ cat(A).
The key point for the reverse inequality is that if cat(Q[X]) ≤ m is finite then H0(A) = Q/(∂(X1)) =
(Q/(∂(X1))) is Artinian, and hence so is H0(A) = Q/(∂(X1)), which means H0(A) is already complete.
It follows that A → A is a quasi-isomorphism. This is because H∗(A) ∼= H∗(A) ⊗Q Q ∼= H∗(Q[Y ])
since each Hi(Q[Y ]) is a finitely generated H0(A) = H0(A) module. Next, since Q/(∂(X1)) is Artinian
there is a system of parameters x for Q which lies in the kernel of Q → S/mS. We can find a subset
z ⊆ X1Q such that ∂(z) = x. Since Q is Cohen-Macaulay x is automatically a regular sequence, so the
homomorphism Q[z] → Q/(x) is a quasi-isomorphism. Note that the composition Q → Q → Q/(x) is
surjective since Q/(x) is Artinian.
Chapter 3. Koszul Duality in Local Commutative Algebra 64
Now, if A → A[Y ]'−→ A/mm+1
A is a minimal model, and α : A[Y ] → A is a splitting, then we can
build the following diagram
A A[Y ]
A
A/(x, z).
' α
β
'
Since the composition A → A → A/(x, z) is a surjective quasi-isomorphism, an A algebra lift β exists
as shown in the diagram, by lemma 5. This shows that cat(A) ≤ cat(A) and concludes the proof.
3.3 The Mapping Theorem
The purpose of this section is to prove the following theorem, which has been discussed already in the
previous section.
Theorem 20. If φ : A→ B is a surjective homomorphism of minimal models then
catA ≥ catB.
Remark 15. Suppose that φ : A→ B is a morphism of minimal models such that φ∗ : π∗(A)→ π∗(B)
is surjective. Then gr(A) → gr(B) is surjective, and it follows that φ : A → B is surjective. According
to the theorem cat A ≥ cat B, and by proposition 11 we can conclude that catA ≥ catB. Therefore the
hypotheses of the mapping theorem can be weakened to surjectivity of φ∗ : π∗(A)→ π∗(B).
From now on we take A = Q[X] and B = P [Y ].
Both P and Q are regular local rings, so the kernel of Q → P is generated by a regular sequence
x in mQ r m2Q. Thus the Koszul complex Q → Q[t]
'−→ P with t = Σx is a model for P . We need a
theorem of Avramov and Iyengar which generalises Gulliksen and Schoeller’s theorem 11 from section
2.7, it explains how to extend this Koszul complex to an acyclic closure of φ.
Theorem (Avramov-Iyengar [19, theorem 1.1]). There are sets Y , Z ⊆ Q[X] which together freely
generate Q[X] = Q[Y , Z], such that the kernel of φ is (x, Z) and φ maps Y bijectively onto Y . Moreover
φ admits an acyclic closure of the form
Q[X]→ Q[X, t]〈U〉 '−→ P [Y ]
where U is a graded set with a bijection ΣZ ∼= U, Σz 7→ uz, such that
∂(uz) = z modulo(m2Q[X],mQ[X](U)(1)
)for all z in Z.
We are treating t as a set of free exterior variables, but these are the same as odd divided power
variables, so can freely switch between the two notations Q[X, t]〈U〉 = Q[X]〈t, U〉.
Chapter 3. Koszul Duality in Local Commutative Algebra 65
Since P → P [Y ] is a semi-free extension and P [X]〈U〉 → P [Y ] is a surjective quasi-isomorphism
there is a P algebra splitting α : P [Y ]→ P [X]〈U〉.
Lemma 17. In the above situation, the splitting satisfies α(Y ) ⊆ (mP [X])P [X]〈U〉.
Proof. We use induction on degree. Assume α(Y<n) ⊆ (mP [X])P [X]〈U〉. Our goal is to show that if
y is in Yn then α(y) is in (mP [X])P [X]〈U〉 as well. By minimality ∂(y) ∈ m2P [Y ] so ∂(α(y)) = α(∂(y))
is in (m2P [X])P [X]〈U〉 by assumption. But we will show that this is impossible unless f = α(y) is in
(mP [X])P [X]〈U〉.P [X]〈U〉 has a P module basis of monomials in X and U with divided powers of elements of U ,
and an element is in (mP [X])P [X]〈U〉 precisely when all its monomials which are purely divided power
monomials in U have coefficients from mP . So, if f is not in (mP [X])P [X]〈U〉 we may consider an element
z in Z of maximal degree i such that uz appears in a divided power monomial with unit coefficient in
f . Denote Z ′ = Zi r z. By considering each monomial in f separately, and using the precise descrip-
tion of the differential described in Avramov and Iyengar’s theorem, we see that ∂(f) = zf ′ modulo
(mP , X<n, Z′,m2
P [X])P [X]〈U〉, where f ′ is the sum of the divided power monomials in f involving Σz, each
with the divided power exponent of Σz reduced by 1. But zf ′ is not in (mP , X<n, Z′,m2
P [X])P [X]〈U〉, so
we reach the contradiction that ∂(f) is not in (m2P [X])P [X]〈U〉. This completes the induction step.
Proof of theorem 20. If we assume that cat(Q[X]) = m is finite, then by definition we have a diagram
of local dg algebras
Q[X] C
Q[X]/mm+1Q[X]
'
making Q[X] a retract of C. The only assumption we need to make about C is that x is regular on it,
so that C ⊗Q Q[t]→ (C/x) is a quasi-isomorphism. This is easy to arrange.
We obtain a dg algebra C〈t, U〉 = C⊗Q[X]Q[X]〈t, U〉 and a quasi-isomorphism C〈t, U〉 '−→ (C/x)〈U〉.Consider the two ideals I = (mC/x)(C/x)〈U〉 and J = (mP [X])P [X]〈U〉/(mm+1
P [X]). From these we obtain
subalgebras A = P + I ⊆ (C/x)〈U〉 and B = P + J ⊆ P [X]〈U〉/(mm+1P [X]) (not the same A and B as in
the original statement).
With this setup, we separate the proof into two claims.
Claim 1: cat(A) ≤ m. Claim 2: P [Y ] is a retract of A.
Having established these we will be done by lemma 14.
We start with claim 1. Apply − ⊗Q[X] Q[X]〈t, U〉 to the quasi-isomorphism C'−→ Q[X]/mm+1
Q[X] to
get another one C〈t, U〉 '−→ Q[t,X]〈U〉/(mm+1Q[X]). We have a commutative diagram of dg algebras
C〈t, U〉 Q[t,X]〈U〉/(mm+1Q[X])
(C/x)〈U〉 P [X]〈U〉/(mm+1Q[X]).
'
' '
The left vertical arrow is a quasi-isomorphism by our assumption on C, and the right vertical arrow is
a quasi-isomorphism by lemma 16. Hence the lower arrow is a quasi-isomorphism. Now we have two
Chapter 3. Koszul Duality in Local Commutative Algebra 66
short exact sequences
0 I (C/x)〈U〉 k〈U〉 0
0 J P [X]〈U〉/(mm+1Q[X]) k〈U〉 0.
'
By the five lemma the left-most map I → J is a surjective quasi-isomorphism. It follows that the
homomorphism A → B is also a surjective quasi-isomorphism. Note also that ``(B) ≤ m + 1 since by
construction Jm+1 = 0 . Therefore cat(A) ≤ m by lemma 14.
It remains to establish claim 2. If we apply−⊗Q[X]P [X]〈U〉 to the split monomorphismQ[X]→ C we
obtain a split monomorphism β : P [X]〈U〉 → (C/x)〈U〉. In the discussion before lemma 17 we remarked
that P [Y ] is a P algebra retract of P [X]〈U〉 via α. The composition βα is a split monomorphism as
well, and by lemma 17 we have βα(Y ) ⊆ β((mP [X])P [X]〈U〉) ⊆ (mC/x)(C/x)〈U〉. Therefore βα factors
through the subalgebra A:
P [Y ] (C/x)〈U〉
A
βα
making P [Y ] a retract of A.
Remark 16. The mapping theorem can be seen as a generalisation of the fact that a sub-Lie algebra of
a free Lie algebra is itself free (see [86, proposition A.1.10]). Indeed, suppose that L′ is a sub-algebra of
a free graded Lie algebra L (concentrated in strictly positive degrees). Then one may construct minimal
models A = C(L∨) and B = C(L′∨) with π∗(A) = L and π∗(B) = L′, and a morphism A→ B inducing
the inclusion L′ → L on homotopy Lie algebras. By the argument in the proof of proposition 9 we have
cat(A) ≤ 1. Since L′ → L is injective A → B is surjective. Hence by the mapping theorem cat(B) ≤ 1
and therefore L′ is also free. Note however that Avramov does make use of this fact about Lie algebras
for his characterisation of Golod dg algebras in [11].
Example 7 (Large homomorphisms). Levin [88] defined a surjective homomorphism φ : (R, k)→ (S, k)
to be large if the induced map TorR∗ (k, k) → TorS∗ (k, k) is surjective. This is closely related to the
hypothesis of the mapping theorem. By theorem 3.4 in [21] (which appears with a simpler proof as
corollary 4 below), φ is large precisely if π∗(R) → π∗(S) is surjective. Taking compatible minimal
models Q[X]'−→ R and P [Y ]
'−→ S we see that R → S is large if and only the induced homomorphism
Q[X]→ P [Y ] is surjective. Hence by the mapping theorem cati(R) ≥ cati(S) for all i.
In general, given a sequence of local homomorphisms ψ : A→ R and φ : R → S one might say that
φ is large relative to A if φ∗ : π∗(ψ)→ π∗(φψ) is surjective. In this case the mapping theorem says that
cati(ψ) ≥ cati(φψ) for all i.
The following corollary strengthens [21, theorem 5.6]. It is a consequence of the mapping theorem
and another theorem of Avramov and Iyengar.
Corollary 3. If φ : R→ S is an almost small homomorphism then cat1(φ) ≤ edim(R)− depth(S) and
cat2(φ) ≤ edim(S)− depth(S).
Chapter 3. Koszul Duality in Local Commutative Algebra 67
Theorem (Avramov-Iyengar [21, theorem 4.11]). Let φ : R → S be a surjective local map. Assume R
has a minimal regular presentation Q→ R, and choose a minimal set x ⊆ Q such that P = Q/x→ S is
a minimal regular presentation. Let Q[t] be the Koszul complex with t = Σx resolving P over Q. If φ is
almost small, one can build the following commutative diagram in which every row, column and diagonal
is a minimal model:Q Q[t] P
Q[X] Q[t,X, T ] P [X,Y ]
R R[t, Y ] S.
'
'
'
' ' '
'
In [21] minimality of the diagonal is not stated, but the proof explicitly shows that ∂(t,X, Y ) ⊆(mQ) + (t,X, Y )2.
Proof of corollary 3. Let R → R′ → S be a minimal Cohen factorisation. By [21, proposition 4.8] the
surjective homomorphism R′ → S is almost small.
In the notation of the theorem applied to R′ → S, we have a surjection k[t,X, Y ]→ k[t, Y ] of minimal
semi-free dg algebras over k. By the mapping theorem and proposition 7
cat1(φ) = cat k[t, Y ] ≤ cat k[t,X, Y ] ≤ sup(k[t,X, Y ]) = pdQ(S)
and pdQ(S) = edim(R′) − depth(S) = edim(R) + edim(R′/mR′) − depth(S), with edim(R′/mR′) =
edim(S/mS) by minimality of the Cohen factorisation. Similarly
cat2(φ) = cat k[Y ]/(Y1) ≤ cat k[X,Y ] ≤ sup(k[X,Y ]) = pdP (S)
and pdP (S) = edim(S)− depth(S).
3.4 Functoriality of the Tate model for the diagonal
The main result of this section is theorem 21, which concerns functoriality of the Tate model for the
diagonal.
In this section Q is a local ring and Q[X] is a semi-free Q algebra with X in positive degrees. Later
on we will take Q to be regular, but we make no such assumption now. The Tate model for the diagonal
was constructed in theorem 10. We state a slightly simplified version here:
Theorem. Let Q → Q[X] be a semi-free extension of a local ring (Q, k) with X degree-wise finite and
concentrated in strictly positive degrees. The multiplication map Q[X,X] → Q[X] has a Tate model of
the form
Q[X,X]→ Q[X,X]〈ΣX〉 → Q[X].
The differential satisfies ∂(Σx) = 1⊗ x− x⊗ 1− Σ∂1(x) modulo decomposables in m2Q[X,X]〈ΣX〉 + (ΣX)(2).
If moreover Q → Q[X] is minimal as a semi-free extension then Q[X,X]〈ΣX〉 is minimal as a dg
module over Q[X,X].
Chapter 3. Koszul Duality in Local Commutative Algebra 68
Recall that in remark 1 it was briefly indicated that the Tate model for the diagonal can be constructed
using twisting cochains, and from this perspective its form is a reflection of the Poincare-Birkhoff-Witt
theorem. Details on that construction will appear elsewhere.
Example 8. Let (R, k) be a local ring with a minimal regular presentation Q → R, and extend this
to a minimal model Q[X] → R. Let X0 denote some choice of minimal generating set for mQ. One
may also build an acyclic closure R → R〈U〉 → k of the residue field of R. The deviations of R are
by definition the numbers εn(R) = card(Un). Recall from section 3.1 that Avramov has shown that we
have card(Xn−1) = εn(R) for all n ≥ 1. A proof was sketched in section section 3.1. But one may also
deduce this from theorem 10, which produces a minimal Tate model Q[X,X]→ Q[X,X]〈ΣX〉 → Q[X].
First, if K = Q〈ΣX0〉 is the Koszul complex with ∂(ΣX0) = X0, then the lifting property 5 applied to
Q[X] → k'←− K provides a dg algebra morphism Q[X] → K. One may then apply − ⊗Q[X] K to
the Tate model for the diagonal to obtain quasi-isomorphisms Q[X]〈ΣX0, ΣX〉'−→ K
'−→ k, which by
adjunction gives a quasi-isomorphism R〈ΣX0, ΣX〉'−→ k. This is an acyclic closure which is witness to
Avramov’s equalities.
Theorem 21. Let Q be a complete local ring and let ξ : Q[X]→ Q[Y ] be a homomorphism of minimal
semi-free Q algebras. Both Q[X] and Q[Y ] have Tate models as in theorem 10, and ξ extends to a
homomorphism of divided power algebras, as in the commutative diagram below
Q[X,X] Q[Y, Y ]
Q[X,X]〈ΣX〉 Q[Y, Y ]〈ΣY 〉
Q[X] Q[Y ].
ξ⊗ξ
ξ
ξ
Moreover we can build ξ so that ξ(Σx) = Σξ1(x) modulo decomposables m2Q[Y,Y ]〈ΣY 〉 + (ΣY )(2), for x ∈ X.
In the statement of the theorem ξ1 : XQ → Y Q is the strict part of ξ. That is, the part which
preserves the weight decomposition Q[X] =⊕
w SymwQ(X) and Q[Y ] =
⊕w Symw
Q(Y ).
Remark 17. The analogous theorem for a morphism of semi-free dg algebras Q[X] → P [Y ] over a
homomorphism Q→ P follows from theorem 21 by adjunction.
Let us introduce some notation to use in the next lemma. We will denote by K the kernel of
Q[X,X]〈ΣX〉 → Q[X], and K(n) is the ideal generated by (x ⊗ 1 − 1 ⊗ x) for all x ∈ X≤n, and
all the divided powers (ΣX)(1). Note that K =⋃K(n). Because of the form of the differential in
the Tate model for the diagonal we have for any x ∈ Xn that ∂(Σx) = x ⊗ 1 − 1 ⊗ x + q where
q ∈ K(n − 1) + mQK(n). Below we will also use the notation Dec = m2Q[X,X]〈ΣX〉 + (ΣX)(2) for
decomposables, and sDec = (X ⊗ 1, 1⊗X, ΣX)2 + mQ(ΣX)(1) + (ΣX)(2) for strongly decomposables.
Lemma 18.
1. Any element t in Dec ∩K can be written as t = ∂(u) + v with u ∈ mQ(ΣX)(1) and v ∈ sDec.
2. Any element f in K can be written as f = ∂(g) + h with both g, h ∈ (ΣX)(1).
Chapter 3. Koszul Duality in Local Commutative Algebra 69
Proof. For 1, assume that t has degree n. Then we can write t =∑x∈Xn a1(x ⊗ 1 − 1 ⊗ x) + v0 =∑
(a1∂(Σx) + a1t1) + v0 with a1 ∈ mQ and t1 ∈ Dec and v0 ∈ sDec (we are suppressing the indexes in
these summations). I will show by induction that for all i we can write
t =∑
(a1∂(Σx1) + ...+ ai∂(Σxi) + aiti + v0 + a1v1 + ...+ ai−1vi−1)
where aj ∈ miQ and tj ∈ Dec ∩ K and vj ∈ sDec and all the xi are in Xn. The base case has been
dealt with. Next one writes ti =∑a(x ⊗ 1 − 1 ⊗ x) + vi =
∑(a∂(Σx) + ati+1) + vi with a ∈ mQ and
ti+1 ∈ Dec and vi ∈ sDec as before. Substituting this into the previous expression for t finishes the
induction step. Part 1 is then established by setting u = a1Σx1 + a2Σx2 + ... and v = v0 + a1v1 + ... using
completeness of Q.
For 2 it suffices to show the following instead: Any element f in K can be written as f = ∂(g) + h
with g ∈ (ΣX)(1) and h ∈ (ΣX)(1) + mQK. Using completeness of Q a similar induction to part 1 will
finish the proof.
Towards this we take f ∈ K(n), and proceed by induction on n. So write f = r + s with r ∈(X≤n ⊗ 1− 1⊗X≤n) and s ∈ (ΣX)(1). It suffices to prove the lemma for r.
Write r =∑x∈X≤n(x ⊗ 1 − 1 ⊗ x)a =
∑∂(Σx)a −
∑qa, where the q ∈ K(n − 1) + mQK(n) are as
in the paragraph before the lemma, and a ∈ Q[X,X]〈ΣX〉. The first summation differs from∑
(Σx)∂(a)
by ∂(∑
(Σx)a), while the second summation is in K(n− 1) + mQK, so we are done by induction.
Proof of theorem 21. We just need to define ξ on ΣX, which we do inductively. So assume we already
have ξ<n : Q[X,X]〈ΣX<n〉 → Q[Y, Y ]〈ΣY<n〉. Fix x ∈ Xn. By lemmas 6 and 7, in order to define ξ(Σx)
we must find an element y ∈ (ΣY )(1)n whose boundary is ξ<n(∂(Σx)), and which has the form y = Σξ1(x)
modulo Dec.
Recall that ∂(Σx) = x⊗ 1− 1⊗ x modulo Dec. This means ξ<n(∂(Σx)) = ξ1(x)⊗ 1− 1⊗ ξ1(x) + s =
∂(Σξ1(x)) + t, where s and t are in Dec.
Since t is a decomposable element in the kernel of Q[Y, Y ]〈ΣY 〉 → Q[Y ] we can write t = ∂(u) + v as
in part 1 of lemma 18. Note that v is automatically in the kernel K of Q[Y<n, Y<n]〈ΣY<n〉 → Q[Y<n].
Since t is a cycle so v. But K is acyclic so v = ∂(f) for some f in K. We apply part 2 of the lemma to get
v = ∂(f − ∂(g)) with h = f − ∂(g) in (ΣY<n)(1). Having degree n+ 1, h is automatically decomposable.
Thus y = Σξ1(x) + u+ h has the desired properties, and we set ξ(x) = y.
The Functoriality of Avramov and Iyengar
In [21] Avramov and Iyengar state a theorem which they call a functorial enhancement of the equalities
card(Xn−1) = εn(R). Their proof is extremely technical and spans more than six pages. Theorem 21
will allow us to give a shorter proof of their result, which we state after some setup.
Let φ : (R, k) → (S, l) be a local homomorphism of complete local rings. Then R and S admit
compatible, minimal Cohen presentations as in the diagram below
Q P
R S.
Chapter 3. Koszul Duality in Local Commutative Algebra 70
By the lifting property of lemma 5 one may extend this to a morphism of minimal models
Q P
Q[X] P [Y ]
R S.
'
ξ
'
And hence a morphism of minimal semifree dg algebras φsf : l[X] → l[Y ]. At the same time, by a
standard lifting property analogous to proposition 5 there is a morphism of acyclic closures
R S
R〈U〉 S〈V 〉
k l.
' '
And hence a morphism of free divided power algebras
TorR(k, l) TorS(l, l)
l〈U〉 l〈V 〉.φγ
Corollary 4 (Avramov-Iyengar [21, Theorem 3.4]). In the above above situation, for every n ≥ 2 there
is a a commutative diagram of homomorphisms of l vector spaces
indn−1(l[X]) indn−1(l[Y ])
indγn(l〈U〉) indγn(l〈V 〉)
indn−1(φsf)
∼= ∼=indγn(φγ)
in which the vertical maps are isomorphisms.
Let ξ : P [X,X]〈ΣX〉 → Q[Y, Y ]〈ΣY 〉 be the morphism of Tate models for the diagonal given by
Remark 17. By uniqueness, we are free to use the construction of acyclic closures explained in example
8. So let s and t be minimal generating sets for mP and mQ respectively. Proposition 5 produces a
commutative diagram of dg algebras
Q[X] P [Y ]
Q〈Σt〉 P 〈Σt〉
k l.
' '
From this one easily produces a morphism R〈Σt, ΣX〉 → S〈Σs, ΣY 〉 of acyclic closures constructed as in
example 8. It is now clear from the description ξ(Σx) = Σξ1(x) modulo decomposables in theorem 21
Chapter 3. Koszul Duality in Local Commutative Algebra 71
that the following diagram commutes
indn−1(l[X]) lXn−1 lYn−1 indn−1(l[Y ])
indγn(l〈t, ΣX〉) lΣXn lΣYn indγn(l〈s, ΣY 〉),
thus we have deduced Avramov-Iyengar’s theorem from theorem 21.
Finally, we state a more general version of Avramov-Iyengar’s theorem, which can be established
using the same proof. This version can be applied to the homotopy Lie algebra of a homomorphism,
and in particular we needed it to establish theorem 16.
Let (P,m, k) and (Q, n, l) be complete regular local rings, and let Q→ A = Q[X] and P → B = Q[Y ]
be minimal semi-free extensions (we do not assume A or B are absolutely minimal). If we take a
minimal generating set a for m and b for n then we can build acyclic closures A → A〈Σa, ΣX〉 '−→ k and
B → B〈Σb, ΣY 〉 '−→ l using theorem 10.
Theorem 22. In the notation above, a local homomorphism ξ : A → B induces a uniquely defined
homomorphism Torξ : TorA∗ (k, l) → TorB∗ (l, l) of divided power Hopf algebras over l, and there is a
commutative diagram
H∗(ΣmA/m2A)⊗k l H∗(ΣmB/m
2B)
indγTorA∗ (k, l) indγTorB∗ (l, l).
Σind(ξ)
∼= ∼=
indγ(Torξ)
Non-minimality makes little difference since ΣmA/m2A can only have a differential at the first step.
We include this because it was necessary for theorem 16.
The theorem follows from the existence of a homomorphism ξ : A〈Σa, ΣX〉 → A〈Σa, ΣX〉 → B〈Σb, ΣY 〉extending ξ, with the property that ξ(Σx) = Σξ(x) modulo decomposables. This is produced by theorem
21.
3.5 Some Long Exact Sequences
In this section we discuss long exact sequences for the homotopy Lie algebra (or coalgebra). In short, to
a sequence (Q, u)ψ−→ (R, k)
φ−→ (S, l) of local homomorphisms one would like to associate a Jacobi-Zariski
exact sequence
· · · → πi+1(φ)→ πi(ψ)⊗k l→ πi(φψ)→ πi(φ)→ πi−1(ψ)⊗k l→ · · ·
This is always possible when the residual characteristic is zero. In general however it is necessary
to impose additional hypotheses on φ and ψ. We will introduce a variation π∗(−) which produces an
analogous long exact sequence in all situations. Then we discuss the question of calculating this variation,
and in particular of when π∗(−) agrees with π∗(−).
Chapter 3. Koszul Duality in Local Commutative Algebra 72
Gulliksen Minimality and Avramov’s Six-Term Exact Sequences
We first explain how Avramov established the existence of a long exact sequence of homotopy Lie algebras
(in some circumstances) by building on work of Gulliksen. This subsection is essentially classical, if
somewhat revisionist.
We begin with some definitions abstracted from [64, section 3]. We say that a semi-free extension
A → A[X] is Gulliksen minimal if ∂(A[X]even) ⊆ mAB + m2A[X]. With divided powers the condition is
reversed: a Tate extension A→ A〈U〉 is called Gulliksen minimal if ∂(A〈U〉odd) ⊆ mAB + (U)(2).
Theorem 23 (Gulliksen [64, lemma 3.2.1 and theorem 3.2.3]). Let R → R〈U〉 be a Gulliksen minimal
Tate extension of a local ring R. There is a minimal Tate extension R → R〈V 〉 and a surjective
quasi-isomorphism R〈U〉 '−→ R〈V 〉 of divided power algebras which induces a quasi-isomorphism of
indecomposables
indγ(R〈U〉) '−→ indγ(R〈V 〉).
Essentially the same proof shows the following semi-free version:
Theorem 24. Let R → R[X] be a Gulliksen minimal semi-free extension of a local ring R. There is
a minimal semi-free extension R → R[Y ] and a surjective quasi-isomorphism R[X]'−→ R[Y ] of divided
power algebras which induces a quasi-isomorphism of indecomposables
mR[X]/m2R[X]
'−→ mR[Y ]/m2R[Y ].
Let φ : (R, k) → (S, l) be a local homomorphism, and build a minimal model Q[X]'−→ R. At this
point we can take a minimal Cohen factorisation of the Q → Q′ → S of the composition Q → R → S.
Note that Q′ is automatically regular. Then we apply proposition 4 to the surjection Q′[X] → S to
obtain a minimal Cohen model Q[X]→ Q′[X,Z]'−→ S. Crucially, Q′[X,Z] can and typically does fail to
be absolutely minimal, since the differential ∂(z) of z ∈ Z may involve indecomposables from Q′[X]. In
any case, tensoring down produces a minimal model R→ R′[Z] = R⊗Q[X]Q[X,Z]'−→ S. The following
diagram summarises the situation
Q′[X,Z]
Q[X] R′[Z]
R S.
'
' 'φ
In particular, we obtain a fibre sequence of semi-free dg algebras
Q[X] −→ Q′[X,Z] −→ Q[Z]
where Q[Z] = k⊗Q[X]Q′[X,Z] = k⊗RR′[Z]. By taking indecomposables we get a short exact sequence
of chain complexes
0 −→mQ[X]
m2Q[X]
⊗k l −→mQ′[X,Z]
m2Q′[X,Z]
−→mQ[Z]
m2Q[Z]
−→ 0. (3.1)
Theorem 25 (Avramov). If wcat(φ) is finite then the extension Q′ → Q′[X,Z] is Gulliksen minimal.
A published proof of this is difficult to find. However, the case that φ is flat is in [7, theorem 1.1],
and the proof there can be adapted to general case.
Chapter 3. Koszul Duality in Local Commutative Algebra 73
Having established Gulliksen’s condition, one constructs Avramov’s long exact sequence as follows.
It follows from theorem 24 that there is a minimal semi-free extension Q′ → Q′[Y ′] and surjective
quasi-isomorphisms
Q′[X,Z]'−−→ Q′[Y ′]
'−−→ S,
with the property that mQ′[X,Z]/m2Q′[X,Z] → mQ′[Y ′]/m
2Q′[Y ′] is a quasi-isomorphism as well. If the
homomorphism Q′ → S fails to be a minimal Cohen presentation then we may find a regular sequence
q ⊆ mQ′ r m2Q′ such that P = Q′/q → S is. After possibly making a change of basis, there is a
subset y ⊆ Y ′1 with ∂(y) = q. We set Y = Y ′ r y, and then we have a minimal Cohen model P [Y ] =
Q′[Y ′]/(q, y)'−→ S and it is easy to see that the induced map of indecomposables
mQ′[Y ′]
m2Q′[Y ′]
'−−→mP [Y ]
m2P [Y ]
is a quasi-isomorphism as well (it is in isomorphism except in degrees 1 and 2, where we collapse a con-
tractible subcomplex). Putting all this together, there is a natural isomorphism ΣH∗(mQ′[X,Z]/m2Q′[X,Z])
∼=Σ(mP [Y ]/m
2P [Y ])
∼= π∗(S).
Observe that the Gulliksen minimality condition implies that the differential of mQ′[X,Z]/m2Q′[X,Z]
vanishes on even elements. Therefore half of the connecting homomorphisms vanish in the long exact
homology sequence associated to 3.1. After applying Σ this looks like
0 π2i(R)⊗k l π2i(S) π2i(φ)
π2i−1(R)⊗k l π2i−1(S) π2i−1(φ) 0.δ
(3.2)
These are the six-term exact sequences of Avramov.
Remark 18. When char(k) = 0 the conclusion of theorem 24 holds without any minimality hypothesis
(this follows from [64, lemma 3.2.1]). In other words, taking indecomposables is an exact functor in
characteristic zero. It follows that these six term exact sequences exist without any additional hypotheses
on φ.
How one can patch these sequences together into a Jacobi-Zariski long exact sequence of relative
homotopy Lie coalgebras is dealt with by proposition 12 and theorem 30.
Translating Between Semi-Free Models And Tate Models
The above sketch is historically somewhat apocryphal, since Avramov and Gulliksen both worked on
the divided power side. We will describe now how the two approaches are directly connected.
First we recall that if R → R〈U〉 '−→ k is an acyclic closure then there is a canonical isomorphsim
πi(R) ∼= indγ(k〈U〉)i; see the discussion in section 3.1 or corollary 4, or one may deduce this from
theorem 10 as we do now.
First recall the construction of the dg algebra Q′[X,Y ] as above, starting from the local homomor-
phism φ : R→ S. According to theorem 10 we have a Tate model for the diagonal
Q′[X,Z,X,Z]〈ΣX, ΣZ〉 '−−→ Q′[X,Z]'−−→ R′[Z]
'−−→ S.
Chapter 3. Koszul Duality in Local Commutative Algebra 74
Let Q′[t]'−→ l be the Koszul complex resolving l over the regular ring Q′. The standard lifting property
of lemma 5 produces a Q′ algebra map Q′[X,Z] → Q′[t]. Now we can apply Q′[t]⊗Q′[X,Z] − to obtain
a quasi-isomorphism Q′[t,X,Z]〈ΣX, ΣZ〉 '−→ Q′[t]'−→ l, and by adjunction we get a quasi-isomorphism
S〈t, ΣX, ΣZ〉 '−→ l.
This is a Tate model for S → l. By the description of the differential in theorem 10 there is a canonical
isomorphism
Σ
(mQ′[X,Z]
m2Q′[X,Z]
)∼= indγ(l〈t, ΣX, ΣZ〉).
Therefore the semi-free extension Q′ → Q′[X,Z] is Gulliksen minimal if and only if the Tate extension
S → S〈t, ΣX, ΣZ〉 is Gulliksen minimal. This explains the shift between the two definitions of Gulliksen
minimality.
We use the same construction to build an acyclic closure R → R〈s, ΣX〉 '−→ k, starting from the
Koszul complex Q[s]'−→ k. We use theorem 21 (and remark 17) to produce a morphism of divided power
algebras ξ : Q[X,X]〈ΣX〉 → Q′[X,Z,X,Z]〈ΣX, ΣZ〉 compatible with the inclusion Q[X]→ Q′[X,Z] such
that ξ(Σx) = Σx for x ∈ X. This is enough to witness the Tate extension S → S〈t, ΣX, ΣZ〉 as the
composition of two minimal Tate extensions
S → S〈s, ΣX〉 → S〈s, r, ΣX, ΣZ〉
with t = s ∪ r. The result is a fibre sequence of divided power algebras
l〈s, ΣX〉 → l〈s, r, ΣX, ΣZ〉 → l〈r, ΣZ〉.
Much as in the semi-free case one obtains a short exact sequence of indecomposables, and from here one
can hope to obtain a long exact sequence of homotopy Lie coalgebras. Avramov argues directly in [7,
theorem 1.1] (in the case that φ is flat) that l〈s, r, ΣX, ΣZ〉 is Gulliksen minimal, and obtains his six-term
exact sequences by these means.
We have shown that by going through the Tate model for the diagonal one can translate between
semi-free models and Tate models, with Gulliksen minimality being equivalent on either side. Both sides
have advantages, but it seems that the semi-free side is more intuitive.
The Six-Term Exact Sequences for πγ∗(φ)
When φ : R → S is surjective one can take an acyclic closure R → R〈U〉 '−→ S. Then we can take the
fibre k〈U〉 = k ⊗R R〈U〉 ' k ⊗LR S. Thinking of this as a divided power analogue of the minimal model
for S over R, we define
πγ∗(φ) = Σindγk〈U〉
The uniqueness of acyclic closures makes πγ∗(φ) well-defined. Certainly nothing like theorem 15 holds
for this object, but the following considerations make πγ∗(φ) a useful computational tool.
We continue R〈U〉 to an acyclic closure R〈U〉 → R〈U, V 〉 '−→ k. Base changing reveals an acyclic
closure S〈V 〉 '−→ k. In order to compare πγ∗(φ) with π∗(S) and π∗(R) we need to understand the extent
to which the Tate model R→ R〈U, V 〉 '−→ k can fail to be minimal. Avramov remarks that his proof of
Chapter 3. Koszul Duality in Local Commutative Algebra 75
[7, theorem 1.1] shows the following theorem.
Theorem 26 (Avramov [7, remark 1.10]). If R → S admits an acyclic closure which is minimal as a
complex of R modules, then R→ R〈U, V 〉 is Gulliksen minimal. Therefore from the short exact sequence
0 −→ indγ(k〈U〉) −→ indγ(k〈U, V 〉) −→ indγ(k〈V 〉) −→ 0
we obtain a long exact homology sequence which splits into six-term exact sequences
0 πγ2i+1(φ) π2i(R) π2i(S)
πγ2i(φ) π2i−1(R) π2i−1(S) 0.δ
(3.3)
The theorem applies in particular to large homomorphisms by [88, theorem 2.5] (large homomor-
phisms were independently proven to have this property by Avramov and Rahbar-Rochandel, according
to loc. cit.)
Example 9. Recall from section 2.7 that a surjective homomorphism φ : R → S with kernel I is a
quasi-complete intersection if its two-step Tate model is acyclic. In other words, it admits an acyclic
closure R → R〈U1, U2〉'−→ S. In this case minimality forces R〈U1, U2〉 to be minimal as a complex,
therefore theorem 26 applies in this situation. It is clear that πγ1(φ) = 0, πγ2(φ) ∼= kU1∼= k ⊗S I/I2 and
πγ3(φ) ∼= kU2∼= k ⊗S H1 where H1 = H1(R〈U1〉) is the first Koszul homology of I (see [31]). In fact, it
follows readily from the definitions that quasi-complete intersections are characterised by the vanishing
of πγ≥4(φ), see theorem 33 for a variation on this. We obtain an exact sequence
0 −→ k ⊗S H1φ2−−−→ π2(R) −→ π2(S) −→ k ⊗S I/I2 −→ π1(R)
φ1−−−→ π1(S) −→ 0,
and isomorphisms φi : πi(R) ∼= πi(S) for i ≥ 3. This appears as [18, theorem 5.3]. The case of a map of
embedded deformations (which are natural examples of qci homomorphisms, see [18]) appears already
in [7, remark 1.10] (but note that qci homomorphisms had not been defined at the time of op. cit.)
The lifting property of lemma 5 means that a minimal model and an acyclic closure can always be
compared:
R R〈U〉
R[X] S.
'
'
Passing to fibres produces a homomorphism k[X]→ k〈U〉, and then taking indecomposables builds the
comparison map π∗(φ)→ πγ(φ). This map will be an isomorphism in characteristic zero (by uniqueness
of minimal models) or when φ is complete intersection, But in general it will fail dramatically to be so.
Below we factor this comparison through another variation π∗(φ) which has the advantage of producing
exact sequences in all situations (and the disadvantage of being slightly more mysterious).
The Mixed Version of Gulliksen’s Theorem
Now we state a mixed version of the two theorems 23 and 24. It could easily be expanded to include
these theorems as special cases, but for notational convenience we refrain from doing so.
Chapter 3. Koszul Duality in Local Commutative Algebra 76
Theorem 27. Let R → R[X] be a minimal semi-free extension of a local ring (R, k), and let R[X] →R[X]〈U〉 be a minimal Tate extension. There is a semi-free mixed extension R→ R[Y ]〈V 〉 such that
1. the extension is minimal in the sense that the differential of indγ(k[Y ]〈V 〉) vanishes;
2. there is a surjective quasi-isomorphism φ : R[X]〈U〉 '−→ R[Y ]〈V 〉 of dg R algebras with divided
powers on (U)(1) and (V )(1) respectively;
3. and φ induces a quasi-isomorphism
indγ(k ⊗R φ) : indγ(k[X]〈U〉) '−−→ indγ(k[Y ]〈V 〉).
Our notation could be ambiguous, so let us emphasise that we do not claim that the reduced mixed
extension R → R[Y ]〈V 〉 can be obtained as a composition R → R[Y ] → R[Y ]〈V 〉 (i.e. ∂(Y ) could
involve terms from W ). The proof doesn’t show this, and we don’t use it below.
We need the following lemma, which is a direct consequence of lemma 8.
Lemma 19. Let R[X]〈U〉 be a mixed extension of a local ring (R, k) and suppose X ′ → R[X]〈U〉 and
U ′ → (U)(1) are two graded functions from positively graded sets. Then the induced map R[X ′]〈U ′〉 →R[X]〈U〉 is an isomorphism if and only if the induced function
X ′ ∪ U ′ → indγ(k[X]〈U〉)
makes X ′ ∪ U ′ a basis of the k vector space indγ(k[X]〈U〉).
Proof of theorem 27. As a graded vector space indγ(k[X]〈U〉) ∼= kX + kU . Since the two extensions
k → k[X] and k[X] → k[X]〈U〉 are minimal the differential looks like kXindγ(∂)←−−−−− kU (i.e. all other
components of indγ(∂) vanish with respect to this decomposition). Let U ′ ⊆ kU be a basis for the kernel
of indγ(∂) : kU → kX and let V ⊆ kU be a basis for any k linear complement to kU ′. Choose a lift
f : U ′ ∪ V → (U)(1) and set X ′ = ∂(f(U ′)) ⊆ R[X]〈U〉. Note that the image of X ′ in indγ(k[X]〈U〉)is precisely indγ(∂)(U ′) ⊆ kX. Finally, choose a subset Y ⊆ kX which freely spans a complement to
k(indγ(∂)(U ′)), and a lift g : Y → R[X]〈U〉.Together f and g induce a homomorphism of graded algebras R[X ′, Y ]〈U ′, V 〉 → R[X]〈U〉 (forgetting
the differential or now). By lemma 19 this homomorphism is an isomorphism.
It follows that the composition
φ : R[Y ]〈V 〉 → R[X ′, Y ]〈U ′, V 〉 → R[X]〈U〉 → R[X]〈U〉/(X ′, (f(U))(1))
is an isomorphism as well. But now note that by construction the ideal (X ′, (f(U))(1)) is closed under the
differential of R[X]〈U〉. In this way R[Y ]〈V 〉 obtains a well-defined differential making it a dg algebra
with divided powers supported on (V )(1), and there is a surjective homomorphism φ : R[X]〈U〉 →R[Y ]〈V 〉 of dg algebras with divided powers.
By construction the differential of indγ(k[Y ]〈V 〉) ∼= kY + kV vanishes, so condition 1 is clear.
It is also clear that indγ(k[X]〈U〉)→ indγ(k[Y ]〈V 〉) is a quasi-isomorphism, because by construction
this map kX + kU → kY + kV simply collapses the contractible subcomplex kX ′+ kU ′, so 3 holds also.
It remains to show that φ is a quasi-isomorphism. By the derived Nakayama lemma it suffices to show
that k⊗φ : k[X]〈U〉 → k[Y ]〈V 〉 is a quasi-isomorphism. This is another application of lemma 1. We use
Chapter 3. Koszul Duality in Local Commutative Algebra 77
the weight filtration with k[X]〈U〉(n) =∑i+j=n(X)i(U)(j) and k[Y ]〈V 〉(n) =
∑i+j=n(Y )i(V )(j). After
passing to the associated graded algebras the differential of k[Y ]〈V 〉 vanishes, while gr(k[X]〈U〉) becomes
isomorphic to k[Y ]〈V 〉 ⊗ k[X ′]〈U ′ : ∂(U ′) = X ′〉. It is well-known that k[X ′]〈U ′ : ∂(U ′) = X ′〉 ' k is
acyclic (see for example [44, proposition 1.9]), and part 2 follows.
Definition of π∗(φ) and Long Exact Sequences
In general a local homomorphism will fail dramatically to induce a long exact sequence of homotopy Lie
coalgebras like 3.2 (see for example [9]). In this section we present a modification of the homotopy Lie
coalgebra which will produce a long exact sequence in all situations.
It is well-known that the Andre-Quillen (co)homology functors also repair this defect. By simplicial
methods Andre and Quillen independently associated a graded Lie algebra D∗(φ; l) to a local homomor-
phism φ : (R, k) → (S, l) (and more generally). The excellent formal properties of these functors is the
foundation for many important results in commutative algebra and algebraic geometry.
There is a natural homomorphism of graded Lie algebras D∗(φ; k)→ π∗(φ) (see [2] for the absolute
situation). This comparison map is an isomorphism when char(l) = 0 and always in low degrees (see
[106]).
However D∗(φ; l) can be very difficult to compute, and it often produce strange results from the
perspective of Koszul duality. For instance, if k is the field with two elements and φ is the inclusion
k → k[x, y]/(x, y)2 then the difficult calculation of D∗(φ; k) is performed by Goerss in [56]. On the other
hand π∗(φ) is the free Lie algebra on two elements of degree 1, as predicted by Koszul duality.
The variation π∗(φ) we present in this section shares some of the good properties of D∗(φ, l) while
being much more computable, and while giving the expected Koszul duality results in some important
cases. For example π∗(φ) ∼= π∗(φ) will be a free Lie algebra whenever φ is Golod. Homotopically π∗(φ)
does not appear to have any significant meaning (there is no analogue of theorem 15), but it seems to
be computationally quite useful.
We can give some intuition as follows. π∗(φ) only depends on the derived fibre k⊗LR S, and somehow
this doesn’t seem to be enough data to extract a long exact sequence (except in characteristic zero). We
will construct a mixed model for k ⊗LR S, and π∗(φ) will be given by the (shifted) indecomposables of
this mixed model. The way the semi-free and divided power structures are made to interact uses slightly
more information about the homomorphism φ, and results in a long exact sequence in all situations.
Roughly, we will see that π∗(φ) interpolates between π∗(φ) and πγ∗(φ).
All this said, we start with an extremely simple definition of π∗(−) which may make it appear unlikely
to be useful. If φ : (R, k)→ (S, l) is a local homomorphism then we set
π∗(φ) = H∗(cone(φ∗ : π∗(R)⊗k l→ π∗(S))).
One can equally well define π∗(φ) = π∗(φ)∨, but we will work homologically in this section. Almost
tautologically there is a long exact sequence
· · · → πi+1(φ)→ πi(R)⊗k l→ πi(S)→ πi(φ)→ πi−1(R)⊗k l→ · · · (3.4)
associated to any local homomorphism. We can slightly generalise this as follows:
Chapter 3. Koszul Duality in Local Commutative Algebra 78
Proposition 12. Let (Q, u)ψ−→ (R, k)
φ−→ (S, l) be a sequence of local homomorphisms. There is a long
exact sequence
· · · → πi+1(φ)→ πi(ψ)⊗k l→ πi(φψ)→ πi(φ)→ πi−1(ψ)⊗k l→ · · ·
Proof. This is an immediate consequence of the Octahedral axiom (which certainly holds for complexes
of l vector spaces). As stated for example in [95], it says that the two unraveled triangles (long exact
sequences from 3.4) below can be interlaced with a third, indicated by the dashed arrows
π∗(Q)⊗u l π∗(S) π∗(φ) Σπ∗(ψ)⊗k l
· · · π∗(R)⊗k l π∗(φψ) Σπ∗(R)⊗k l · · ·
Σ−1π∗(φ) π∗(ψ)⊗k l Σπ∗(Q)⊗u l Σπ∗(S).
The real interest of this proposition comes from the fact that there is a natural map π∗(φ)→ π∗(φ),
and that this map is often an isomorphism. This is our strategy for uncovering Jacobi-Zariski exact
sequences for the homotopy Lie algebra. By these means π∗(φ) is a potentially useful computational
tool.
An Alternative Construction for π∗(φ)
As before we start with a local homomorphism φ : (R, k) → (S, l). The key to calculating π∗(φ) is an
alternative construction which explains how it interpolates between π∗(φ) and πγ∗(φ) (see also theorem
30 specifically).
We show in particular that now there is a natural model F for the (completed) fibre k ⊗LRS such
that indγ(F ) = π∗(φ). As above we repeatedly use the Tate model for diagonal to mediate between free
extensions and divided powers.
As before we take a minimal Cohen model Q[X]'−→ R. Then using a minimal Cohen factorisation
Q → Q′ → S we obtain a minimal Cohen presentation P = Q′/q → S (the regular sequence q is as
above, in the construction of Avramov’s six-term exact sequences). Then we extend to a minimal model
P → P [Y ]'−→ S. The basic lifting property of lemma 5 produces a diagram
Q[X] P [Y ]
R S
φ
' '
φ
which can be used to define φ∗ : π∗(R) → π∗(S). Similarly, if Q[t]'−→ k is the Koszul complex on a
Chapter 3. Koszul Duality in Local Commutative Algebra 79
minimal set x of generators for mQ, then there is a homomorphism of Q algebras Q[X] → Q[t] lifting
the quotient homomorphism Q[X]→ k.
We use theorem 10 to build Q[X,X]〈ΣX〉 '−→ Q[X] with ∂(Σx) = x modulo decomposables, for x ∈ X.
From this we apply Q[t]⊗Q[X] − to obtain a quasi-isomorphism Q[t,X]〈ΣX〉 '−→ Q[t], and finally we use
φ to apply −⊗Q[X] P [Y ] and build P [Y ]〈t, ΣX〉.Note that through Q[X]
'−→ R there are natural quasi-isomorphsims Q[t,X]〈ΣX〉 '−→ R〈t, ΣX〉 '−→ k.
Through P [Y ]'−→ S is also a natural quasi-isomorphism P [Y ]〈t, ΣX〉 → S〈t, ΣX〉 ∼= S⊗R R〈t, ΣX〉. Hence
the dg algebra P [Y ]〈t, ΣX〉 is a model for the derived tensor product S ⊗LRk.
Finally, we examine the differential of P [Y ]〈t, ΣX〉. Firstly ∂(t) is determined by the image of x
under Q → P . Next, for x ∈ X we have ∂(Σx) = φ(x) modulo decomposable elements, thanks to the
description of the differential in theorem 10. Putting all these facts together one obtains the following
Theorem 28. Let (R, k) → (S, l) be a local homomorphism. The dg algebra P [Y ]〈t, ΣX〉 constructed
above is naturally a model for S ⊗LRk. There is a canonical isomorphism
Σ indγ(P [Y ]〈t, ΣX〉) ∼= Σ
((mPm2P
+ lY )indγ(∂)←−−−−−− (l ⊗k
mQm2Q
+ lX)
)∼= cone(φ∗ : π∗(R)⊗k l→ π∗(S)),
and therefore π∗(φ) ∼= ΣH∗(indγ(P [Y ]〈t, ΣX〉)).
We use this description of π∗(φ) to build the comparison maps in the following theorem. In it we
continue to use the notation from the constructions made above. Note also that by base-changing the
minimal Cohen factorisation Q → Q′ → S along Q → R we obtain a minimal Cohen factorisation
R→ R′ → S. We extend this to a minimal Cohen model R→ R′[Z]'−→ S. The fibre k⊗QQ′ ∼= k⊗RR
′
will be denoted Q. The dg algebra Q[Z] is a minimal model and by definition π∗(φ) = π∗(Q[Z]).
Theorem 29. There are quasi-isomorphisms of dg algebras
Q[Z]'←−− Q′[t, Z]
'−−→ P [Y ]〈t, ΣX〉
which induce the identity map on TorR∗ (S, k). Taking indecomposables produces a comparison map
π∗(φ) −→ π∗(φ).
Assume that φ is surjective and take an acyclic closure R〈U〉 '−→ S. There is a quasi-isomorphism
P [Y ]〈t, ΣX〉 '−−→ k〈U〉
of dg algebras with divided powers supported on (t, ΣX)(1) and (U)(1) respectively, which again induces
the identity on TorR∗ (S, k). Taking indecomposables produces a comparison map
π∗(φ) −→ πγ∗(φ).
Proof. We may take a minimal model Q′[X]→ Q′[X,Z]'−→ S and then obtain R′[Z] as R⊗Q[X]Q
′[X,Z].
Recall that Q[t]'−→ k is the Koszul complex of Q and that we have a homomorphism Q[X] → Q[t]
of Q algebras. It extends to a homomorphism Q′[X] → Q′[t]'−→ k ⊗Q Q′. Note that there is a natural
Chapter 3. Koszul Duality in Local Commutative Algebra 80
quasi-isomorphism
Q′[t, Z] = Q′[t]⊗Q′[X] Q′[X,Z]
'−→ (k ⊗Q Q′)⊗Q′[X] Q′[X,Z] = Q[Z].
This constructs the first quasi-isomorphism in the theorem. It is easy to see that this homomorphism
induces a quasi-isomorphism upon taking indecomposables:
mQ′[t,Z]
m2Q′[t,Z]
'−−→mQ[Z]
m2Q[Z]
.
Because this map is an isomorphism except in degrees 0 and 1 where it collapses the contractible
subcomplex l · (x)∂←− l · (t).
Intermediate to the construction of the mixed model P [Y ]〈t, ΣX〉 we can build the dg algebra
P [X,Y ]〈ΣX〉 = Q[X,X]〈ΣX〉 ⊗Q[X] P [Y ]. It comes with a homomorphism from Q′[X] (using the re-
maining copy of X on the left and the surjection Q′ → P ) and a natural surjective quasi-isomorphism
P [X,Y ]〈ΣX〉 '−→ S.
Now lemma 5 produces a lift
Q′[X] P [X,Y ]〈ΣX〉
Q′[X,Z] S.
'
'
(3.5)
Applying Q′[t] ⊗Q′[X] − to this lift produces the desired map Q′[t, Z] → P [Y ]〈t, ΣX〉. It is a quasi-
isomorphism because Q′[X,Z] is semi-free as a dg Q′[X] module and P [X,Y ] is semi-free as a dg P [X]
module. Taking indecomposables and shifting produces the desired map
π∗(φ) = Σ
(mQ[Z]
m2Q[Z]
)∼=←− ΣH∗
(mQ′[t,Z]
m2Q′[t,Z]
)−→ ΣH∗(indγ(P [Y ]〈t, ΣX〉)) ∼= π∗(φ).
It remains to establish the second part of the theorem, in which we assume that φ is surjective. We
continue with the same notation as before, but now we may assume that Q = Q′ and k = l.
We complete our acyclic closure to obtain R→ R〈U〉 '−→ S.
We can build a minimal model Q → Q[s, Y ]'−→ S lifting the model P → P [Y ]
'−→ S. Here s is a
degree one set of variables with ∂(s) = q, so that Q[s] ' P .
Similarly to above the Tate model for the diagonal builds a surjective quasi-isomorphismQ[X, s, Y ]〈ΣX〉 '−→S. We now have a commutative diagram
Q R〈U〉
Q[X, s, Y ]〈ΣX〉 S.
'
'
α
There are no problems in using the analogue of lemma 5 for this mixed extension, since R〈U〉 has divided
powers supported on the ideal of all positive degree elements. Therefore a homomorphism of α dg Q
algebras with divided powers exists, as indacted by the dashed arrows. We tensor this homomorphism
Chapter 3. Koszul Duality in Local Commutative Algebra 81
with the Koszul resolution β : Q[t]'−→ k to obtain a quasi-issomorphism
Q[t, s, Y ]〈ΣX〉 = Q[t]⊗Q[X] Q[X, s, Y ]〈ΣX〉β⊗Q[X]α−−−−−−→ k ⊗Q[X] R〈U〉 = k〈U〉
Finally, this descends along the quasi-isomorphism Q[t, s, Y ]〈ΣX〉 '−→ P [t, Y ]〈ΣX〉 to give the desired
homomorphism, as stated in the theorem.
An Application to Lie∞ Algebras
Before moving on we briefly point out that in characteristic zero theorem 28 actually recovers one of the
main results in a paper of Fiorenza and Manetti [50]. Our proof is simpler because it entirely avoids the
perturbation methods of op. cit. The reader may wish to skip this small subsection.
Some background of Lie∞ algebras is given at the end of section 4.2. We will only say that a Lie∞
algebra structure on a degree-wise finite dimensional graded vector space L is specified by a differential
making k[Σ−1L∨] into a semi-free dg algebra. A morphism L → L′ of Lie∞ algebras is a morphism
k[Σ−1L′∨]→ k[Σ−1L∨] of semi-free dg algebras.
Corollary 5. Let k be a field of characteristic zero and let χ : L→ L′ be a morphism of cohomologically-
positively graded Lie∞ algebras over k. The shifted cone Σ−1cone(χ) naturally admits the structure of an
Lie∞ algebra in such a way that the projection Σ−1cone(χ)→ L naturally extends to a Lie∞ morphism.
The connectedness assumptions are unnecessary, we include them only because of our conventions
on semi-free dg algebras. The constructions go through without issue in the general situation.
The point of this result is that even when L and L′ are simply dg Lie algebras, the shifted cone
Σ−1cone(χ) does not necessarily admit a dg Lie algebra structure. That is, higher operations must be
taken into account to make this construction. Fiorenza and Manetti also prove that the Maurer-Cartan
functor associated to Σ−1cone(χ) controls the deformations of χ.
Proof. The dual notion of a Lie∞ coalgebra structure on L is simply a differential making k[ΣL] a semi-
freee dg algebra. We establish the dual result: to a morphism χ : L′ → L of Lie∞ coalgebras there is a
Lie∞ coalgebra structure on cone(χ) such that the inclusion L→ cone(χ) is one of Lie∞ coalgebras.
In terms of semi-free dg algebras, we are given k[Σ−1L′]→ k[Σ−1L] and we must construct an extension
k[Σ−1L] → k[Σ−1L,L′] such that the linear part of the differential of k[Σ−1L,L′] recovers the shifted
cone Σ−1cone(χ). Since the characteristic of k is zero there is non-difference between k[Σ−1L,L′] and
k[Σ−1L]〈L′〉 and this was exactly the construction we made in theorem 29 (in the proof of that theorem
we assumed that k[Σ−1L] and k[Σ−1L′] were minimal, but this assumption was not used).
Properties of π∗(φ)
Theorem 30. Let φ : R→ S be a local homomorphism. We can identify π∗(φ) in the following cases:
1. If wcat(φ) is finite then the comparison π∗(φ)→ π∗(φ) is an isomorphism. This applies for example
if φ has finite flat dimension, or is Golod, or more generally when φ is almost small.
2. If φ is surjective and admits an acyclic closure R → R〈U〉 '−→ S which is minimal as a complex
then π∗(φ) → πγ∗(φ) is an isomorphism. This applies for example if φ is large or quasi-complete
intersection.
Chapter 3. Koszul Duality in Local Commutative Algebra 82
3. If the residual characteristic of S is zero then π∗(φ)→ π∗(φ)→ πγ∗(φ) are both isomorphisms.
Proof. Part 3 follows from remark 18 and the fact that all algebras have divided powers in characteristic
zero. Part 1 uses Avramov’s six term exact sequences 3.2 while part 2 uses the divided power version
3.3. Since the proofs are similar we only treat 1.
The point is that the dg algebra morphisms from theorem 29 allow us to construct a commutative
diagram of dg algebras
Q[X] Q′[X,Z] Q′[Z, t]
Q[X] P [Y ] P [Y ]〈t, ΣX〉.
' '
Here we have used the notation from theorem 29. Then we take indecomposables, shift and pass to
homology. By theorems 24, 25 and 28 this results in two long exact sequences
· · · πi(R)⊗k l πi(S) πi(φ) · · ·
· · · πi(R)⊗k l πi(S) π∗(φ) · · ·
We conclude that the comparison π∗(φ)→ π∗(φ) is an isomorphism using the five lemma.
Combining this theorem with proposition 12 produces a Jacobi-Zariski long exact sequence for the
homotopy Lie algebra (or for πγ∗(−)) in a variety of situations.
Example 10. Suppose that the local homomorphisms (Q, u)ψ−→ (R, k)
φ−→ (S, l) are both Golod. By [8,
theorem 3.5] they are also small, so their composition is small, and therefore all three of φ, ψ and φψ
have finite 0-category by corollary 3. By theorem 30 and (the linear dual of) proposition 12 there is a
long exact sequence
· · · → πi+1(φ)← πi(ψ)⊗k l← πi(φψ)← πi(φ)← πi−1(ψ)⊗k l← · · ·
sandwiching π∗(φψ) between the two free Lie algebras π∗(φ) and π∗(ψ).
We can generalise the situation for small and large homomorphisms (and give a shorter, independent
proof that does not use the long exact sequences of Avramov or the Gulliksen minimality conditions).
Theorem 31. Let (R, k)→ (S, l) be a local homomorphism.
1. If φi : πi(R)⊗k l→ πi(S) is injective for all even i then π∗(φ)→ π∗(φ) is an isomorphism.
2. If φ is surjective and φi : πi(R) ⊗k l → πi(S) is surjective for all odd i then π∗(φ) → πγ∗(φ) is an
isomorphism.
This is quite surprising. Only the injectivity of φeven is enough to guarantee that we have the full
six term exact sequences 3.2. And only the surjectivity of φodd is enough to guarantee that we have the
full six term exact sequences 3.3.
Proof. We start with 1. We use the constructions and notation of theorem 29 and its proof.
Recall that the homomorphism Q′ → S factors through the minimal Cohen presentation Q′ → P =
Q/q → S. Recall that we also constructed a quasi-isomorphism Q′[X] → R′ and a semi-free model
Chapter 3. Koszul Duality in Local Commutative Algebra 83
Q′[X]→ Q′[X,Z]'−→ S. Since edim(Q′) = edim(R′) it follows (possibly making a change of basis) that
we can assume Z = Z ′ ∪ v for some v ⊆ Z1 with ∂(v) = q.
We need to slightly improve the construction in the proof of theorem 29: when we built the lift in
diagram 3.5 we could have first passed to the quotient Q′[X,Z]'−→ P [X,Z ′] = Q′[X,Z]/(q, v) and done
this via a lift P [X,Z ′]→ P [X,Y ]〈ΣX〉.The upshot is that we may assume that the homomorphism Q′[t, Z]→ P [Y ]〈t, ΣX〉 of that theorem
factors as
Q′[t, Z]'−→ P [t, Z ′]
'−→ P [Y ]〈t, ΣX〉.
Having made this technical assumption, part 1 is now a relatively painless consequence of theorem 27.
That theorem says we can reduce P [Y ]〈t, ΣX〉 to a quasi-isomorphic mixed extension P [W ]〈V 〉 which is
minimal in the sense that indγ(l[W ]〈V 〉) has vanishing differential, and moreover the induces chain map
indγ(l[Y ]〈t, ΣX〉)→ indγ(l[W ]〈V 〉) is a quasi-isomorphism:
P [t, Z ′] P [Y ]〈t, ΣX〉
P [W ]〈V 〉.
'
α '
The key point is that the hypothesis that φi : πi(R)⊗k l→ πi(S) is injective for even i implies that the
construction in theorem 27 must remove all the even parts of ΣX. In other words, V is concentrated in
odd degrees. Therefore P [W ]〈V 〉 = P [W,V ] is actually semi-free. Since indγ(l[W ]〈V 〉) = ind(l[W,V ])
has no differential the extension P → P [W,V ] is minimal. Therefore α : P [t, Z ′] → P [W,V ] is a quasi-
isomorphism or minimal semi-free P algebras. By [12, lemma 7.2.3] α is an isomorphism. In particular
it induces an isomorphism on indecomposables. We have canonical isomorphisms
π∗(φ) ∼= ΣH∗
(mP [t,Z′]
m2P [t,Z′]
)∼= ΣH∗
(mP [W,V ]
m2P [W,V ]
)∼= ΣH∗(indγ(P [Y ]〈t, ΣX〉)) ∼= π∗(φ).
This establishes part 1.
The proof of part 2 is similar (and simpler), using theorem 27 to produce a model for P [Y ]〈t, ΣX〉which involves only divided power variables. We omit it.
Besides it being striking that injectivity of φeven is equivalent to having the full six term exact
sequences 3.2, the theorem is useful because this condition is clearly stable under composition.
Corollary 6. Let (Q, u)ψ−→ (R, k)
φ−→ (S, l) be two local homomorphisms such that cat2(ψ) and cat2(φ)
are both finite. Then there is a Jacobi-Zariski long exact sequence of homotopy Lie coalgebras
· · · → πi+1(φ)→ πi(ψ)⊗k l→ πi(φψ)→ πi(φ)→ πi−1(ψ)⊗k l→ · · ·
Proof. Since wcat2(ψ) and wcat2(φ) are both finite we know that ψ2i : π2i(Q) ⊗u k → π2i(R) and
φ2i : π2i(R)⊗k l→ π2i(S) are all injective by Avramov’s six term exaxct sequences (3.2). It follows that
(φψ)2i : π2i(Q) ⊗u l → π2i(S) are injective as well. By part 1 of theorem 31 there are isomorphisms
π∗(ψ) ∼= π∗(ψ) and π∗(φ) ∼= π∗(φ) and π∗(φψ) ∼= π∗(φψ). After this the desired long exact sequence
comes from proposition 12.
Chapter 3. Koszul Duality in Local Commutative Algebra 84
In the theorem below parts 1 and 2 are true and well-known for the homotopy Lie algebra, while
part 3 fails except in characteristic zero. One of the purposes of π∗(−) is to extend a step further to 3.
Theorem 32. A local map φ : R→ S is
1. weakly regular if and only if π≥2(φ) = 0;
2. complete intersection if and only if π≥3(φ) = 0;
3. quasi-complete intersection if and only if π≥4(φ) = 0.
Since the interesting part of theorem 32 is 3 we separate it and elaborate slightly in the following
theorem. All three parts are consequences of theorem 27, so we omit the proofs of 1 and 2. One direction
of the theorem below is due to Avramov, Henriques and Sega [18, theorem 5.3] (this was discussed in
example 9), we establish the converse.
Theorem 33. If φ : (R,m, k)→ (S, n, l) is a local homomorphism then the following are equivalent:
1. φ is quasi-complete intersection;
2. π≥4(φ) = 0;
3. π≥3(R)⊗k l→ π≥3(S) is an isomorphism.
And in this case there is an exact sequence in low degrees
0→ π3(φ)→ π2(R)⊗k l→ π2(S)→ π2(φ)→ π1(R)⊗k l→ π1(S)→ π1(φ)→ 0.
Moreover if R→ R′ → S is a minimal Cohen factorisation then
π1(φ) = (n/mS)⊗S l, π2(φ) = (I/I2)⊗S l and π3(R) = H1(KI)⊗S l
where I is the kernel of R′ → S and KI is the Koszul complex on a minimal generating set for I.
Because of long exact sequence ... → π∗(R) → π∗(S) → π∗(φ) → ... the condition π≥4(φ) = 0 is
equivalent to πi(R) → πi(S) being an isomorphism for i ≥ 4 and π3(φ) being injective. Note that
exactly same statement holds for Andre-Quillen homology, by the characterisation of Blanco, Majadas
and Rodicio [31]. Thus qci homomorphisms behave as if they were characterised by vanishing of π∗(−), if
only there were a long exact sequence of homotopy Lie algebras. This is exactly the role that π∗(−) plays.
Part 3 of the theorem says that in this situation π3(R) ⊗k l → π3(S) is automatically an isomorphism
as well.
Proof. Let R → R′φ′−→ S be a minimal Cohen factorisation of φ. By the Jacobi-Zariski exact sequence
(proposition 12) associated to this factorisation, and part 1 of theorem 32, all of the conditions 1, 2 and
3 are unchanged if we replace φ with φ′. Therefore we may assume that φ is surjective.
The implications 1 ⇒ 3 ⇒ 2 are all clear from example 9 and the basic long exact sequence ... →π∗(R)→ π∗(S)→ π∗(φ)→ ... So we must show 2 ⇒ 1.
We apply theorem 27 to a dg algebra similar to the one constucted in theorem 28. Let Q[X]'−→ R
be a minimal model, and also build a minimal model Q → Q[Y ]'−→ S (we do not worry about the
fact that Q → S may not be minimal, and so Y1 may be “too large”). Lemma 5 produces a lift
Chapter 3. Koszul Duality in Local Commutative Algebra 85
Q[X] → Q[Y ] of dg Q algebras. We use theorem 10 to build a model Q[X,X]〈ΣX〉 '−→ Q[X], then
finally we apply R ⊗Q[X] − ⊗Q[X] Q[Y ] to build a surjective quasi-isomorphism R[Y ]〈ΣX〉 '−→ S. In
low degrees we were lazy about minimality, but the description of the differential in theorem 10 allows
us to identify indγ(R[Y ]〈ΣX〉)≥2∼= Σ
−1cone(π≥2(R) → π≥3(S)). Therefore the hypothesis 2 implies
that indγ(P [Y ]〈t, ΣX〉) has homology in degrees 1 and 2 only. Theorem 27 produces a surjective quasi-
isomorphism R[Y ]〈ΣX〉 '−→ R[Z]〈U〉 with Z and U both concentrated in degrees 1 and 2 only.
In characteristic zero R[Z]〈U〉 = R〈Z,U〉 and the augmentation R〈Z,U〉 '−→ S is witness to that
fact that φ is qci. If char(k) = p is positive we must show that Z2 = ∅. For degree reasons R[Z]〈U〉 is
automatically minimal as a complex, therefore k[Z]〈U〉 ∼= TorR∗ (S, k). But if z ∈ Z2 this contradicts the
fact that TorR∗ (S, k) has a natural divided power algebra structure supported on the ideal of positive
degree elements, since this means zp 6= 0 should be impossible. Therefore the two step Tate model
R〈Z1, U〉'−→ S is witness to that fact that φ is qci.
Finally, remembering again that φ might not be surjective, we must establish that there is a low degree
exact sequence as shown. There is always such an exact sequence using π∗(φ) in place of π∗(φ). The
calculation that πi(φ) = πi(φ) is as described for i = 1, 2, 3 is relatively simple, and already essentially
part of [18, theorem 5.3] (where surjectivity is assumed), so we leave it to the reader. Details can be
found in the proof of loc. cit. and in example 9 above.
Finally, the conjectures stated by Quillen in [106] for Andre-Quillen homology (and extended by
Avramov in [13]) have an obvious analogue for π∗(−). If φ : (R, k) → (S, l) is a local homomorphism
then πi(φ) = 0 for large i if and only if φi : πi(R)⊗k l→ πi(S) is an isomorphism for large i. So we can
ask:
Question. If there is an n such that φi : πi(R) ⊗k l → πi(S) is an isomorphism for all i > n then
must φ be a quasi-complete intersection?
Chapter 4
The Characteristic Action of
Hochschild Cohomology
This chapter contains some highlights from the preprint [34]. An overview of its contents was given in
the introduction. Everything here is joint work with Vincent Gelinas, and our contribution to this work
should be considered equal.
4.1 Enriching the Characteristic Action
This section is adapted from [34, section 3.1].
Let A be a dg algebra which is augmented over a field k1. This section concerns the characteristic
action of Hochschild cohomology on the derived category, which is a homomorphism χ : HH∗(A,A) →Z(D(A)) (the graded centre Z(D(A)) was defined in section 2.1. This map is important in many ar-
eas of mathematics. In representation theory alone, it has been used in connection with the classical
cohomology operators of Gulliksen, the theory of support varieties and with the obstruction theory for
deforming modules, to name a few applications. As such, it is important to understand the image of the
components χM : HH∗(A,A) → Z(Ext∗A(M,M)). Towards this, we will show that the homomorphisms
χM can be enriched to land in HH∗(R,R), where R = RHomA(M,M). This implies that an endomor-
phism in the image of χM must satisfy some strong homotopy centrality conditions. We also show that
for certain dg modules the image of χM is exactly the image of Π : HH∗(R,R)→ Z(Ext∗A(M,M)).
As explained in section 2.1 Ddg(A) is dg category of all semi-free dg modules over A. It’s homotopy
category D(A) = H0(Ddg(A)) is the full derived category of A. We use the same notation for the graded
category D(A) = H∗(Ddg(A)). Similarly Ddgperf(A) is our notation for the dg category of perfect semi-free
dg modules, and perf(A) = H0(Ddgperf(A)) is the perfect derived category of A.
It is well known that the restriction C∗unr(Ddgperf(A), Ddg
perf(A)) → C∗unr(A,A) is a quasi-isomorphism
(this is easily proven with Keller’s theorem in [75], for example). Less well known, it follows from the
work of Toen [118] that the restriction C∗unr(Ddg(A), Ddg(A)) → C∗unr(A,A) is a quasi-isomorphism (see
also [81, section 4.3]). This section contains a proof of this fact. The main theorem is
1The augmentation is not important for this section because we could work with the unreduced bar construction BunrAinstead of BA. Since we are most interested in applying this section to the augmentation module kA we will continue tomake this assumption anyway.
86
Chapter 4. The Characteristic Action of Hochschild Cohomology 87
Theorem 34. The restriction map C∗unr(Ddg(A), Ddg(A))→ C∗unr(A,A) is a quasi-isomorphism. Through
the induced isomorphism in cohomology, the projection map from HH∗(DdgA,DdgA) and the character-
istic action of HH∗(A,A) coincide. That is, the following diagram commutes
HH∗(DdgA,DdgA) HH∗(A,A)
Z(D(A)).
∼=res
Π χ
Given a morphism ζ : A → ΣnA in D(Ae) we get a family of morphisms M ⊗L
A ζ : M → ΣnM
in D(A) which define a natural transformation 1D(A) → Σn. This determines a map χ : HH∗(A,A) ∼=
Ext∗Ae(A,A)→ Z(D(A)), which is a homomorphism of graded k algebras. We call this the characteristic
action of HH∗(A,A) on D(A). The two most important components are for the modules A and k:
HH∗(A,A)
H(A) Z(D(A)) ExtA(k, k).
χA χk
res res
It is not difficult to check that χA and χk are the projection and shearing morphisms respectively, which
were introduced in section 2.5. In that section χk was denoted by χ alone, in this section we reserve χ
for the full characteristic action.
Let us be more explicit about this construction. A Hochschild cocycle ξ : BA → A canonically lifts
to a bilinear map AξA : A⊗π BA⊗π A→ A, and for any dg module M we get an A-linear map
MξA = M ⊗A AξA : M ⊗π BA⊗π A −→M.
Since M⊗πBA⊗πA is a semi-free resolution of M , this MξA represents an element of ExtA(M,M). Nat-
urality up to homotopy is easy to see, so the family MξA defines an element of Z(D(A)). This explicit
description suggests a further enhancement of the characteristic action. First, note that Homπ(M ⊗π
BA,M) ∼= HomA(M ⊗πBA⊗πA,M) has a convolution product given by
φ ^ ψ : M ⊗BA→M ⊗BA⊗BA ψ⊗BA−−−−−→M ⊗BA φ−−→M,
for φ, ψ : M ⊗BA→M . With this product the natural inclusion HomA(M,M)→ Homπ(M ⊗πBA,M)
becomes a quasi-isomorphism of dg algebras (for semi-free M). In fact, this extends to a convolution
enhancement Dconv(A) of the derived category D(A), with NDconv(A)M = Homπ(M ⊗πBA,N). The
obvious embedding Ddg(A)→ Dconv(A) is then a quasi-equivalence of pre-triangulated dg categories.
The point is that the characteristic action χM now lifts to a homomorphism of dg algebras
C∗(A,A)→ Homπ(M ⊗πBA,M) ξ : BA→ A 7→ Mξ : M ⊗BA M⊗ξ−−−−→M ⊗A→M.
In fact it lifts further, all the way to the Hochschild cochain complex of RHomA(M,M), as long as one
is willing to work with two enhancements at once. More generally, one can do this for any set of objects
together.
Chapter 4. The Characteristic Action of Hochschild Cohomology 88
Proposition 13. Let M be a set of objects in D(A) and let R = REndA(M) be the full dg subcategory
of Ddg(A) on M . Denote also by Rconv the corresponding full dg subcategory of Dconv(A). The char-
acteristic action of HH∗(A,A) on H∗(R) lifts canonically to map χ of dg algebras, as in the following
commutative diagram:
C∗unr(A,A) C∗unr(R,Rconv) C∗unr(R,R) HH∗(A,A) HH∗(R,R)
Z(Rconv) Z(R) Z(H∗(R)).
χ
χ Π
'
Πχ
χ
Π
'
In the proposition Z(R) denotes the ‘pre-centre’∏m∈M REndA(m), and similarly for Rconv.
The lift χ takes a cochain ξ : BA → A to the composition BunrRη−→ kM
α−→ Rconv, where α takes
1m to (−) · ξ in Homπ(m ⊗π BA,m). It is not difficult to verify that this is an anti-homomorphism of
dg algebras (i.e. reversing the order of composition, but in particular respecting the differential), but
we can make things clearer by writing them in a more symmetrical way. We can think of M as an
R−A bimodule. Just rewriting things through the tensor-hom adjunction, the top row in the diagram
of proposition 13 becomes2
Homπ(BA,A)→ Homπ(BunrR⊗π M ⊗π BA,M)← Homπ(BunrR,R)
ξ 7→ η · (−) · ξ ζ · (−) · η ←[ ζ
where η is being used for the counits of both BA and BunrR.
This symmetry suggests that the homologically balanced condition from section 2.1 may play a role.
Theorem 35. Let M be a set of objects in D(A) and let R = REndA(M) be the full dg subcategory of
Ddg(A) on M . If the bimodule RMA is homologically balanced then χ : C∗unr(A,A) → C∗unr(R,Rconv) is a
quasi-isomorphism.
In the case that M contains the free module A, the restriction map C∗unr(R,R) → C∗unr(A,A) is a
quasi-isomorphism, inverse in cohomology to χ.
The advantage of the last statement is that the restriction map is one of B∞-algebras (as Keller
points out in [79]), and also that it respects the weight filtration on Hochschild cohomology.
Note that we only need to check half of the definition of homologically balanced: in this context the
map R→ REndA(M) is a quasi-equivalence by definition.
Proof. In light of the above discussion we just need to show that
Φ : Homπ(BA,A)→ Homπ(BunrR⊗πM ⊗πBA,M)
is a quasi-isomorphism. But using the tensor-hom adjunction on the other side we have Homπ(BunrR⊗π
M ⊗πBA,M) ∼= Homπ(BA,Homπ(BunrR ⊗πM,M)) (as usual checking that the various twists match
up). Note though that Homπ(BunrR ⊗πM,M) is simply a convolution model for REndRop(M), that is
Homπ(BunrR⊗πM,M) ∼= HomR(R⊗πBunrR⊗πM,M)← HomR(M,M) is a quasi-isomorphism. Hence by
2The twist on Homπ(BunrR⊗π M ⊗π BA,M) is the one which makes it isomorphic to HomR−A(R⊗π BunrR⊗π M ⊗πBA⊗π A,M).
Chapter 4. The Characteristic Action of Hochschild Cohomology 89
assumption the map Aop → Homπ(BunrR ⊗πM,M) is a quasi-isomorphism, and by lemma 4 this makes
Φ a quasi-isomophism as well.
To make the last statement more precise, we mean that the upper triangle in the following diagram
commutes after taking cohomology
Homπ(BunrR,R) Homπ(BunrR,Rconv)
Homπ(BA,A) Homπ(BA,Aconv) Homπ(BA⊗πA⊗πBA,A).
res
'
res
'
χ
∼=
The outer rectangle clearly commutes. The lower triangle commutes in cohomology because the two
paths are equalised by a quasi-isomorphism, as follows:
Homπ(BA,A) ⇒ Homπ(BA⊗πA⊗πBA,A)∼−→ Homπ(BA,A).
The two maps on the left are the left and right actions ξ 7→ ξ · (−) · η and η · (−) · ξ respectively. The
final map is obtained by pulling back along the natural bicomodule map BA→ BA⊗πA⊗πBA. Either
composition from left to right is the identity. The statement follows.
Now we will deduce theorem 34 by taking M to be the collection of all dg modules in D(A). Those
uneasy about the size of D(A) can just use a sufficiently large set instead, but this subtlety will not be
important here.
Proof of theorem 34. Because of theorem 35, we just need to show that D = DdgA is homologically
balanced as a D −A bimodule. We have a sequence of quasi-fully-faithful functors
Ddgperf(A
op) (Ddgperf(A))op Dop Ddg
perf(Dop),'
RHomAop (−,A)
D⊗LA−
y
where the Yoneda embedding y is quasi-fully-faithful by the Yoneda lemma (note that dg modules in
the image of y are automatically perfect and semi-free). One can check that the diagram commutes (up
to a natural quasi-isomorphism). Now since D ⊗LA − is quasi-fully-faithful it follows from lemma 3 that
Aop → REndDop(D) is a quasi-isomorphism.
Remark 19. The proof works for any dg subcategory between Ddgperf(A) and Ddg(A).
Finally, theorem 35 allows us to give a conceptual proof of the following theorem of Buchweitz, Green,
Snashall and Solberg.
Theorem 36 (Buchweitz-Green-Snashall-Solberg [38]). If A is a Koszul algebra then χk : HH∗(A,A)→Z(ExtA(k, k)) is surjective.
Proof. Since A is Koszul there is a quasi-isomorphism A! ' RHomA(k, k) ' Ext∗A(k, k) respecting the
augmentation module k. By corollary 1 k is homologically balanced as an A! − A bimodule. Therefore
Chapter 4. The Characteristic Action of Hochschild Cohomology 90
by theorem 35 we have a commutative diagram
HH∗(A,A) HH∗(Ext∗A(k, k),Ext∗A(k, k))
Z(Ext∗A(k, k)).
∼=χk
χk Π
Therefore im(χk) = im(Π) = Z(Ext∗A(k, k)).
In the next section we will explain the analogous statement in the non-Koszul case in terms of A∞
centres.
4.2 The A∞ centre
In this section we introduce and begin to study a notion of A∞ centre for minimal A∞ algebras. Let me
remind the reader everything here is joint work with Vincent Gelinas.
To start with, take A to be a strongly augmented dg algebra. From now on we return to the notation
χ = χk and χ = χk. Most of the proof of theorem 36 goes through without any Koszul assumption, and
we get
Corollary 7. If A is a strongly connected augmented dg algebra then there is a canonical isomorphism
making the following diagram commute:
HH∗(A,A) HH∗(A!, A!)
Z(H∗(A!)).
∼=χk
χ Π
In the Koszul case the isomorphism of algebras HH∗(A,A) ∼= HH∗(A!, A!) was first established by
Buchweitz [36]. For dg algebras it is due to Felix, Menichi and Thomas [48], who proved further that it
can be upgraded to an isomorphism of Gerstenhaber algebras.
Our construction is very similar to the one used by Keller in [79], where he recasts things in terms
of the restriction maps (as in theorem 35) to see that the isomorphism lifts to one in the homotopy
category of B∞-algebras. However, the purpose of our approach is to make it clear that the isomorphism
χ interacts well with the characteristic action.
After corollary 7, it remains to understand the image of Π : HH∗(A!, A!) → Z(H∗(A!)). In general
there is no reason for this to be surjective. Instead the answer will be a kind of A∞ centre, which can
be calculated using a minimal A∞ model for A!.
Motivation and Definition of the A∞ Centre
There are several possible notions of centre, and of commutativity, for A∞ algebras, all generalising
the usual concepts for graded algebras. A common answer is that a commutative A∞ algebra is a C∞
algebra (introduced by Kadeishvili in [74]). In characteristic zero certain aspects of commutative algebra
generalise well to C∞ algebras. However, this does not obviously give rise to a notion of centre for A∞
algebras. A more immediate disadvantage is that this property is not even invariant under isomorphisms
of A∞ algebras.
Chapter 4. The Characteristic Action of Hochschild Cohomology 91
Ultimately, E∞ algebras encapsulate the right notion of commutative algebras up to homotopy, as far
as generalising commutative algebra goes. Commutativity at this level is an extra structure on, rather
than a property of, an A∞ algebra. The commutativity property we introduce below is a completely
different and much weaker notion.
Let A be an A∞ algebra. The space of homotopy derivations of A is by definition the positive weight
part of the suspended Hochschild cochain complex
hoDer(A,A) = ΣHomπ(BA,A) = ΣC(1)(A,A).
It is a sub-Lie algebra of ΣC∗(A,A). We can interpret the projection morphism as the obvious map
Π : C∗(A,A) = Σ−1hoDer(A,A) oA −→ A,
just as we saw shearing morphism in terms of coderivations in section 2.5.
The reason for this notation is the following proposition, which goes back to Stasheff and Schlessinger
[108], and ultimately Quillen. A proof is in [69], or more explicitly in [48, Lemma 4.2].
Proposition 14. The dg Lie algebra hoDer(A,A) is quasi-isomorphism invariant, and there is a canon-
ical chain of quasi-isomorphisms
hoDer(A,A)∼−−→ Der(A,A)
of dg Lie algebras, where A = ΩBA. Or more generally, A can be any semi-free dg algebra quasi-
isomorphic to A.
Now, from the short exact sequence
0→ Σ−1hoDer(A,A)→ C∗(A,A)→ A→ 0
we get a connecting homomorphism H∗(A)ad−→ H∗(hoDer(A,A)). In fact, this canonically lifts to the
chain level: the assignment
ad : A −→ hoDer(A,A) ∼=∏n≥1
Σ1−nHom(A
⊗n, A) a 7→
∑n≥1
(−1)|a|Σ1−n[a;−]1,n.
is a chain map giving rise to the above connecting homomorphism in homology3. Here we have used
the higher commutators against a from section 2.5. Under the quasi-isomorphism of proposition 14 this
corresponds to the classical adjoint homomorphism
ad : A −→ der(A,A) a 7→ [a,−].
With this is mind, we define the A∞ centre of a minimal A∞ algebra A to be
Z∞(A) = ker(A
ad−−−→ H∗(hoDer(A,A)))
= im(
HH∗(A,A)Π−−→ A
).
So, a is A∞-central if there exists p ∈ hoDer(A,A) with ad(a) = ∂(p). Writing this out explicitly in
3We were relaxed about signs here because we didn’t specify the isomorphism hoDer(A,A) ∼=∏n≥1 Σ
1−nHom(A⊗n, A).
The map can be written precisely using the shuffle product on BA as in [34, section 3.2]. However, the the signs in (4.1)should be correct, and that expression is what we use to make calculations.
Chapter 4. The Characteristic Action of Hochschild Cohomology 92
terms of the higher multiplications, this means that a is central if for n ≥ 2 the higher commutators
[a;−]1,n : A⊗n → A vanish together up to a sequence of ‘homotopies’ pi : A⊗i → A of degree |a| − i for
i ≥ 1, meaning precisely that [a;−]1,n =∑r+s+t=n
(−1)r(|a|+s)+t(|a|+1)mr+1+s(1⊗r ⊗ ps ⊗ 1⊗t)− (−1)|a|(−1)rs+tpr+1+t(1
⊗r ⊗ms ⊗ 1⊗t). (4.1)
In particular the usual commutator [a;−]1,1 vanishes, and so Z∞(A) ⊆ Z(A). Being the image of a map
of graded algebras, Z∞(A) is itself a graded commutative subalgebra of A.
The intuition is that keeping track of the homotopy p for which ad(a) = ∂(p) gives the full derived
centre, i.e. Hochschild cohomology
HH∗(A,A) = (a, p) : ad(a) = ∂(p) where a ∈ A and p ∈ hoder(A,A)/(boundaries) ,
and forgetting p is exactly the projection morphism.
The reason for introducing Z∞(A) in terms of hoDer(A,A) is that this perspective will be computa-
tionally useful. In particular, examples show that the homotopies pi for the higher commutators in (4.1)
can sometimes be ignored.
Properties of Z∞(A)
If A and A′ are isomorphic minimal A∞ algebras, it follows from naturality of the projection morphism
(diagram 2.1 in section 2.5) that Z∞(A) ∼= Z∞(A′) as graded algebras. In particular, if A is a dg algebra,
then the A∞ centre Z∞(H∗(A)) of a minimal model H∗(A) for A is an invariant of the quasi-isomorphism
type of A. In fact Z∞(H∗(A)) is just the image of Π : HH∗(A,A)→ H∗(A).
If A is formal (isomorphic to a minimal A∞ algebra with vanishing higher structure) then Z∞(A) =
Z(A), but in general A∞ centre will be much smaller than the graded centre.
From theorem 7 we can deduce
Theorem 37. If A is a strongly connected augmented dg algebra then the image of χ : HH∗(A,A) →Ext∗A(k, k) is exactly Z∞(Ext∗A(k, k)).
This is a direct consequence of the definitions and the diagram of theorem 7, along with naturality
(2.1) of the projection morphism from section 2.5). However, it will become useful once we have methods
for computing A∞ centres. Note also that theorem 36 of Buchweitz, Green, Snashall and Solberg is
precisely the special case that A and A! are both formal. It can be generalised to other modules4.
Corollary 8. Let M be a right dg module over A, and take Ext∗A(M,M) to be a minimal A∞ model for
RHomA(M,M). For every dg module M , the characteristic morphism χM : HH∗(A,A)→ ExtA(M,M)
lands in the A∞ centre of Ext∗A(M,M). If M is homologically balanced as an REndA(M)−A bimodule
then im(χM ) = Z∞(Ext∗A(M,M)).
This time the corollary is deduced from proposition 13 and theorem 35. Aside from the augmentation
module k, any generator for Dperf(A) is homologically balanced, for example.
4Technically we have not discussed minimal models for non-augmented dg algebras like RHomA(M,M), but there isno problem with Kadeishivili’s minimal model theorem [73] in this generality. The necessary naturality for Hochschildcohomology is also fine in the non-augmented situation.
Chapter 4. The Characteristic Action of Hochschild Cohomology 93
Remark 20. According to the classical work of Gerstenhaber [52], a first order deformation of a graded
algebra A is classified by a weight 2 element ξ ∈ HH2(A,A). The obstruction to continuing a module M
along such a deformation is Char(ξ) = [MξA] ∈ Ext2A(M,M). The corollary tells us that the obstruction
must actually be in Z∞(ExtA(M,M)). Lowen [91] has extended this picture to deformations of abelian
categories. Here the image of the characteristic morphism contains the obstruction to deforming objects
along a given deformation of an abelian category. This is an important motivation for us to compute
A∞ centres.
In section 4.3 we will give examples of interesting A∞ centres, explaining how one can compute them
with the philosophy of this section. For now, let us note down a few examples with Z∞(A) = A.
We say that a minimal A∞ algebra A is A∞ commutative if Z∞(A) = A, or equivalently if the
projection map Π : HH∗(A,A) → A is surjective. At the end of this section we make note of one
interesting consequence of this condition.
Any algebra with the homotopy type of a commutative dg algebra will be A∞ commutative, but our
condition is much weaker than this.
The shearing map for is split surjective for any Hopf algebra. By theorem 37 this means the Koszul
dual to any strongly connected Hopf algebra is always A∞ commutative.
The algebra C∗(X; k) of cochains on a space X is always A∞ commutative. More generally, E∞
algebras satisfy this condition, because of the next paragraph.
One can see (e.g [53, theorem 7]) from the defining formulas that for any algebra over the “brace
operad”, denoted S2 in [97], the projection morphism is split surjective. Alternatively it is shown in
[124] that the bar construction of such an algebra is a Hopf algebra. It is proven in [96] that this brace
operad S2 is an E2 operad, see also [97]. Since this operad is Σ-split, by [69] any E2 algebra has a model
which is a brace algebra, and hence any E2 algebra is A∞ commutative.
The Associated Lie∞ Algebra and A∞ Commutativity
For the rest of this section we assume that the based field k has characteristic zero. Lie∞ algebras were
discussed tangentially in section 2.8. Here we rapidly record the details we need to establish corollary
9, more information can be found in [90].
To start with we need the symmetric coalgebra and the shuffle product.
Let V be a graded vector space and define Symco(V ) = SymTco(V ) to be the subspace of symmetric
elements with SymTcow (V ) = Sw(V ⊗w). This makes Symco(V ) a sub-coalgebra of Tco(V ). In fact it is the
cofree commutative cocomplete coalgebra on V .
The tensor coalgebra is naturally a Hopf algebra with the shuffle product
(v1 ⊗ ...⊗ vp) ∈ (vp+1 ⊗ ...⊗ vp+q) =∑
σ∈sh(p,q)
(−1)|σ;v|vσ−1(1) ⊗ ...⊗ vσ−1(p+q),
where sh(p, q) is the set of permutations in Sp+q which separately preserve the order of 1, ..., p and
of p + 1, ..., p + q. This is the unique coalgebra morphism m : Tco(V ) ⊗ Tco(V ) → Tco(V ) such that
the corresponding maps mpq : Tcop (V )⊗ Tco
q (V )→ V vanish unless p+ q = 1, in which case our hand is
forced by unitality. With this product Symco(V ) is in fact a sub Hopf algebra of Tco(V ).
Since k has characteristic zero Symco(V ) is exactly the subalgebra of Tco(V ) generated by V under
the shuffle product.
Chapter 4. The Characteristic Action of Hochschild Cohomology 94
Now we define a Lie∞ algebra structure on a graded vector space L to be specified by a degree −1
coderivation b on Symco(ΣL) which is augmented and square zero. By definition the cobar construction
CcoL of L is the dg commutative coalgebra Symco(ΣL) equipped with this differential.
Analogously to the associative case, one can extract from the components bn : Ccon L→ L a sequence
of anti-linear maps ln : L⊗n → L for n ≥ 2. These can be thought of as higher brackets. The condition
that b2 = 0 becomes a sequence of Stasheff-like quadratic identities in the ln generalising the Jacobi
identity. These can be found in [84].
It follows from the discussion in section 2.8 that a graded Lie algebra is precisely a Lie∞ algebra
such that the bar differential b decreases weight by exactly 1. More generally, a dg Lie algebra is a Lie∞
algebra such that b decreases weight by no more than 1. In the other direction, we say that L is minimal
if its bar differential strictly decreases weight. In other words, the underlying differential of L vanishes.
We say that a minimal Lie∞ algebra L is completely abelian if the differential of CcoL vanishes entirely.
Equivalently, all of the operation ln vanish for n ≥ 2.
Lemma 20. Every coderivation p : Tco(V )→ Tco(V ) restricts to Symco(V ).
The lemma follows immediately from the fact that Tco(V ) and Symco(V ) share the same universal
property for coderivations. Or it can be proven by examining Quillen’s formula for p given in lemma 2.
Consequently, if A is an augmented A∞ algebra then bar differential of BA automatically restricts
to Symco(ΣA). This defines a Lie∞ algebra with bar construction CcoALie = (Symco(ΣA), b), which we
call the associated Lie∞ algebra.
Theorem 38. Assume k has characteristic zero. If A is a minimal A∞ algebra which is A∞ commutative
then ALie is completely abelian.
Proof. By induction on weight we show that the bar differential b vanishes on the symmetric tensors
Cco(ALie) ⊆ BA.
Since k has characteristic zero, any weight w + 1 symmetric element can be written as a linear
combination of elements of the form [a] ∈ x, where a is in A and x is symmetric of weight w. Since a is
in the image of the projection map there is a coderivation p which (weakly) decreases weight such that
[b, [a] ∈ −] = [b, p]. Thus
b([a] ∈ x) = bp(x) + (−1)|a|pb(x)− (−1)|a|[a] ∈ b(x).
Since p decreases weight and preserves Cco(ALie) by lemma 20, this formula and the inductive hypotheses
on b means b([a] ∈ x) = 0.
We can rephrase this entirely in terms of dg algebras:
Corollary 9. Assume k has characteristic zero and let A be an augmented dg algebra such that Π :
HH∗(A,A) → H∗(A) is surjective. Then the associated dg Lie algebra ALie is formal and quasi-
isomorphic to an abelian Lie algebra.
The corollary follows by taking a minimal model for A. One only needs to observe that (−)Lie is
functorial even for non-strict morphisms and preserves quasi-isomorphisms.
A dg algebra A satisfying the hypotheses of the corollary need not be quasi-isomorphic to a com-
mutative algebra, so this is an example of what subtler forms of commutativity can be deduced from
A∞-commutativity of the minimal model H∗(A).
Chapter 4. The Characteristic Action of Hochschild Cohomology 95
In the literature these dg Lie algebras have been called quasi-abelian. Understanding when this
condition holds is important because it means the associated deformation problem is extremely simple,
and in particular unobstructed.
A variation on corollary 9 gives a necessary and sufficient condition for deciding when a given Lie
algebra is quasi-abelian.
Corollary 10. Assume k has characteristic zero and let L be a dg Lie algebra over k. Then Π :
HH∗(UL,UL)→ H∗(UL) is surjective if and only if L is quasi-abelian.
This is an immediate consequence of corollary 9 and the following lemma.
Lemma 21. Let L be a dg Lie algebra. If (UL)Lie is quasi-abelian then so is L.
Proof. It follows from the work of Baranovsky [24] that there is a minimal Lie∞ algebra Lmin and a
quasi-isomorphism Lmin '−→ L, and there is a universal enveloping A∞ algebra U(Lmin) which is also
minimal, and a quasi-isomorphism U(Lmin)'−→ UL. Moreover if Lmin happens to be an honest graded
Lie algebra then Baraonovsky’s construction U(Lmin) coincides with the usual universal envelope. And
finally, there is a canonical strict inclusion of Lie∞ algebras Lmin → U(Lmin)Lie.
Now, if (UL)Lie is quasi-abelian then U(Lmin)Lie is completely abelian. But from the strict embedding
Lmin → U(Lmin)Lie it follows that Lmin is completely abelian as well. But then the quasi-isomorphism
Lmin '−→ L means that L is quasi-abelian by definition.
Finally, the following corollary is an algebraic generalisation of a result of Felix, Thomas and Vigue-
Poirrier in rational homotopy theory [49] (see [34, section 5] for more discussion on this).
Corollary 11. Assume k has characteristic zero and let A be a strongly connected dg commutative
algebra such that χ : HH∗(A,A) → Ext∗A(k, k) is surjective. Then A is formal and in fact there is a
quasi-isomorphism Symk(V )'−→ A for some graded vector space V . In other words, A is free up to
homotopy.
Proof. Since A is commutative its bar construction BA is the universal envelope of a unique dg Lie
coalgebra LcoA (see for example [90, section 13.1.10]). Write A! = (BA)∨ = U(LcoA)∨. By corollary 7
the projection Π : HH∗(A!, A!) → H∗(A!) is surjective, so (LcoA)∨ is quasi-abelian by corollary 10. So
there is a quasi-isomorphism L ' (LcoA)∨ of Lie coalgebras, when L is an abelian Lie algebra. Since
k has characteristic zero it follows that there is a quasi-isomorphism Symk(L) ' A!. By theorem 5
A ' A!! ' Symk(L)! ' Symk(Σ−1L∨). By freeness this can be upgraded to a direct quasi-isomorphism
Symk(Σ−1L∨)'−→ A as in the statement.
4.3 Examples and Applications
This section is adapted from [34, section 4]. As such it is also joint work with Vincent Gelinas.
First we admit that, for the sake of writing down more interesting examples, we work here with
a (superficially) more general setup than in the rest of the paper. Instead of working with algebras
augmented over a field k, we work over a semi-simple base k isomorphic to a finite product of copies of
k. Instead of vector spaces we work in this section with (graded or dg) k bimodules on which k acts
Chapter 4. The Characteristic Action of Hochschild Cohomology 96
centrally. It is equivalent to work with (graded or dg) quivers whose set of objects S is indexed by the
factors of k ∼= k×S . Undecorated tensor products and homs are taken over k, and in particular (−)∨ is
the k linear duality Hom(−,k).
An augmented k algebra is by definition a retraction k→ A→ k of k algebras. In particular A is a
(possibly non-symmetric) k bimodule and there is a canonical decomposition of k bimodules A = k+A.
It is equivalent to think of A as a k linear category whose objects are indexed by the factors of k. In
particular the bar construction can be thought of as a cofree cocategory on the same set of objects (as
was hinted at in section 2.1).
We hope the reader accepts in good faith that all the definitions and results of the sections leading
up to this one generalise immediately to this context (simply replace the word ‘algebra’ with ‘category’
throughout).
We need to explain one more notational convention for this section. So far we have not been explicit
about the meaning of the asterisk in HH∗(A,A) or in Ext∗A(k,k), except that we used HH(w)(A,A) to
denote the weight filtration. This also passes to a weight filtration Ext(w)A (k,k) on Ext∗A(k,k) through
the shearing morphism. When A is a graded algebra (that is, when its differential vanishes) the weight
filtration splits into a weight grading. This is also known as the cohomological or resolution degree. Thus
Hochschild cohomology is bigraded with HHw(A,A)i coming from Homπ(BwA,A)i. The same goes for
ExtwA(k,k)i. If x ∈ ExtwA(k,k)i then we will write wt(x) = w and |x| = i.
When A is simply an algebra concentrated in degree zero the internal degree coincides with weight, so
HHw(A,A)i = 0 and ExtwA(k,k)i = 0 unless w+ i = 0. More interestingly, when A is generated in degree
1 as a k algebra, it follows from theorem 7 that A is Koszul if and only if Ext∗A(k,k) is concentrated
along the diagonal according to the bigrading (wt(x),−|x| − wt(x)) (that is, ExtwA(k,k)i 6= 0 implies
2w + i = 0).
For us, interesting and relevant A∞ algebras arise for the most part as the Koszul duals to actual
graded algebras (but this is by no means the only way; for instance, symplectic topology is a notable
source of examples).
The A∞ structure on Kadeishvili’s minimal model A! = Ext∗A(k,k) can be produced by homological
perturbation methods (possibly first exploited by Huebschmann beginning with [71], and made explicit
by Merkulov in [98]). This approach has some theoretical advantages, but it is typically easiest to
produce a minimal model for A! by building the noncommutative Tate model ΩC'−→ A and setting
A! = C∨. This was briefly sketched in section 2.4, but see [25] for details on the construction.
From the Tate model one deduces Keller’s theorem 9. This fundamental calculation is the beginning
of understanding the higher structure on A!. Explicitly understanding the higher structure in full is only
possible in special cases5.
Our main source of examples will be the d-Koszul algebras discussed below. They were introduced
by Berger in [27], and the Koszul dual A∞ structure was computed completely by He and Lu in [66].
Another case where this structure can be understood completely is that of monomial algebras, when
A = kQ/(R) for some quiver Q and some set R of paths in Q. This has been done recently by Tamaroff
in [116] using the noncommutative Tate construction. The same calculation was made independently in
unpublished work of Chuang and King.
5This depends on whether one accepts as explicit the iterative perturbation formulas in [98], or its interpretation interms of trees in [82]. In any case, these formulas depend of a choice of contraction onto Ext∗A(k, k) from some knownmodel for RHomA(k, k).
Chapter 4. The Characteristic Action of Hochschild Cohomology 97
Algebras of finite global dimension
Assume that gldim(A) = g is finite, so that Ext>gA (k,k) = 0. The conditions (4.1) defining a top degree
element ξ ∈ Z∞(ExtgA(k,k)) are vacuous aside from [ξ, s] = 0 for s ∈ Ext0A(k,k) = k, since the non-trivial
higher products against ξ strictly increase weight. The condition [ξ, s] = 0 cuts out the symmetric part
of the k bimodule ExtgA(k,k), which we denote ExtgA(k,k)cyc since these will correspond to cycles in the
quiver path algebra case. So
Proposition 15. If gldim(A) = g is finite then the image of χ : HHg(A,A)→ ExtgA(k,k) is ExtgA(k,k)cyc.
Restricting now to gldim(A) = 2 theorem 9 determines the entire A∞-algebra Ext∗A(k,k), since all
other products vanish for strict unitality or degree reasons. We are left to determine Z∞(Ext1A(k,k)).
Proposition 16. Assume A is finite dimensional with global dimension 2. Then the image of χ :
HH∗(A,A)→ Ext∗A(k,k) is the graded centre Z(Ext∗A(k,k)).
Proof. The result follows from proposition 15 because Z(Ext1(k,k)) = 0 by the No Loop Conjecture
(established in the global dimension 2 case in [59]).
The proposition means we will need to look to algebras of global dimension at least 3 for interesting
examples.
d-Koszul algebras
As we saw is section 2.3 a Koszul algebra is necessarily quadratic. Nevertheless, classical Koszul duality
can be generalised naturally to d-homogeneous algebras. A d-homogeneous algebra is by definition an
algebra with a presentation A = T(V )/(R) for some k bimodule V and some sub-bimodule R ⊆ V ⊗d.If d > 2 then according to theorem 7 the Koszul to such an A cannot be formal. That is, the minimal
model A! = Ext∗A(k,k) must have some non-trivial higher products. Rather, a d-Koszul algebra will be
one whose A∞ structure is as simple as possible subject to the contraints of theorem 9.
These algebras were defined by Berger in [27] in terms of acyclicity of an explicit candidate Koszul
complex. For our definition we can use the following characterisation of Green, Marcos, Martınez-Villa
and Zhang from [58]: a strongly connected d-homogeneous algebra A is d-Koszul if and only if Ext∗A(k,k)
is generated as a k algebra in weight 1 and 2.
All we need to know about them is the structure theorem of He and Lu [66], and the fact, courtesy
of Green, Marcos, Martınez-Villa and Zhang, that our examples are themselves d-Koszul.
First, we mention that some intuition is given by the work Dotsenko and Valette in [41]. They prove
that when A is d-Koszul the A∞ algebra A! satisfies a natural universal property. We sketch this now.
A minimal A∞ coalgebra is called (2, d)-reduced if its higher coproducts ∆n are zero except for
n = 2, d and, if d 6= 2, composing ∆d with itself in any entry is zero.
If we are given the d-homogeneous data (V,R), so V is a k bimodule and R ⊆ V ⊗d, then we can
associate the algebra A = A(V,R) = T(V )/(R).
At the same time, to the d-homogeneous data (W,S) we can associate an A∞ coalgebra as follows.
By definition C = C(W,S) is the universal (2, d)-reduced A∞ coalgebra equipped with a map C → W
such that C∆d−−→ C⊗d →W⊗d factors through S →W⊗d, and, if d 6= 2, such that then the composition
C∆2−−→ C⊗2 →W⊗2 is zero. For any other (2, d)-reduced A∞ coalgebra satisfying these conditions there
is a unique strict morphism C ′ → C over W .
Chapter 4. The Characteristic Action of Hochschild Cohomology 98
There is a twisting cochain C(ΣV, ΣdR) → V → A = A(V,R) and Dotsenko and Valette establish
that this twisting cochain is acyclic if and only if A is d-Koszul. This makes precise the statement that
A! = C(ΣV, ΣdR)∨ should be as simple as possible subject to theorem 9.
By dualising the above universal properties and exhibiting an explicit construction one recovers the
structure theorem of He and Lu, which we state after some setup.
The d-homogeneous dual of A = A(V,R) is by definition the algebra A⊥ = A(V ∨, R⊥), where R⊥ is
the orthogonal complement to R in (V ∨)⊗d ∼= (V ⊗d)∨. The weight w part is denoted Aw⊥.
Theorem 39 (He-Lu [66]). If A is a d-Koszul algebra then there are isomorphisms
Ext2iA (k,k) ∼= Σ
−2iAid⊥ and Ext2i+1A (k,k) ∼= Σ
−2i−1Aid+1⊥ .
The products mn are zero if n 6= 2, d. Under the the above identification m2 is given by
Σ−2iAdi⊥ ⊗ Σ
−2jAdj⊥ Σ−2(i+j)A
d(i+j)⊥
Σ−2i−1Adi+1
⊥ ⊗ Σ−2jAdj⊥ Σ
−2(i+j)−1Ad(i+j)+1⊥
Σ−2iAid⊥ ⊗ Σ
−2j−1Adj+1⊥ Σ
−2nA(i+j)d+1⊥
all of these maps are induced by the standard product on A⊥, after moving all shifts to the front according
to the Koszul sign rule. If d 6= 2 then m2 vanishes on Σ−2i−1Aid+1
⊥ ⊗ Σ−2j−1Adj+1
⊥ . Finally, the operation
md vanishes if any of its inputs have even weight, and on odd weight it is the map
Σ−2i1−1Ai1d+1
⊥ ⊗ · · · ⊗ Σ−2id−1Ajd+1
⊥ Σ−2(i1+...+id+1)A
d(i1+...+id+1)⊥
which is induced from the iterated product m(d)2 on A⊥.
The theorem has some immediate consequences:
Theorem 40. Let A be d-Koszul. Then in even weight the A∞ centre of Ext∗A(k,k) coincides with its
graded centre
Z∞(ExtevenA (k,k)) = Z(Exteven
A (k,k)).
This follows since the higher products vanish on even inputs.
Corollary 12. Let A be d-koszul and assume that A is nilpotent. Then the shearing map χ induces an
isomorphism of graded algebras
HH∗(A,A)/(nil) ∼= Z(Ext∗A(k,k))/(nil)
where (nil) is the nilradical on either side.
Indeed since m(n) : A⊗n → A is 0 for n 0 one easily sees that elements in Homπ(BA,A) are
nilpotent since then f^n = m(n)(f ⊗ ... ⊗ f)∆(n) = 0 for n 0, and thus ker χ consists of nilpotent
Chapter 4. The Characteristic Action of Hochschild Cohomology 99
elements. If d > 2 then odd weight elements in Ext∗A(k,k) are nilpotent, while for d = 2 there is no
higher structure, and the result follows.
Examples
We end by using theorem 39 to calculate a few interesting examples of A∞ centres.
If Q is a quiver with a set of objects Q0 we will take k = k×Q0 . Then the path algebra kQ is by
definition the tensor algebra Tk(Q). The set Qw of paths in Q of length w is a basis for the weight w
part of kQ. We will always place Q in homological degree 1. So weight coincides with internal degree
and all our examples are strongly connected.
When drawing a quiver Q presenting an algebra A = kQ/(R) by a minimal set of relations R ⊆ kQ≥2,
a dashed arrow will indicate a relation connecting two objects. In other words, a dashed arrow from x
to y means that yRx 6= 0.
Our examples will be d-Koszul thanks to the following theorem.
Theorem 41 (Green-Marcos-Martınez-Villa-Zhang [58]). Assume that A = kQ/(R) where R is a set of
paths in Qd. Then A is d-Koszul if and only if R has the d-covering property: for any three composable
paths p, q, r of length at least 1, if both pq ands qr are in R then every subpath of pqr of length d is in
ρ. In particular the d-truncated path algebra kQ/(Qd) is d-Koszul.
The first example is at this point almost classical. It seems that Madsen was the first to compute it
[94].
Example 11. Let A = k[x]/(xd) with |x| = s even and d > 2. Then A! = Ext∗A(k, k) ∼= k[∂x, η] with
|∂x| = −s − 1 and |η| = −sd − 2. Note that as an algebra A! is strictly graded commutative. However
we will show that A! typically isn’t A∞ commutative (depending on the characteristic of k).
The non-trivial higher products are of the form
md(ηi1∂x ⊗ ...⊗ ηid∂x) = η(
∑dm=1 im)+1
Let us calculate the higher commutators. From this formula we see that ad(ηj) = 0, while ad(ηj∂x)(∂⊗d−1x )
is given by
ad(ηj∂x)(∂x, ..., ∂x) =
d−1∑i=0
(−1)i(−1)imd(∂⊗ix ⊗ ηj∂x ⊗ ∂⊗d−1−i
x ) = d · ηj+1
This is 0 if and only if char(k) | d. When char(k) - d this cannot be the image of a coboundary ∂π(p) for
any p ∈ hoDer(A!, A!) = Homπ(BA!, A!). Indeed, writing (Σ∂x)⊗d−1 = [∂x|...|∂x], we have
∂π(p)([∂x|...|∂x]
)= ∂(p)
([∂x|...|∂x]
)+
d−2∑n=1
[π, ..., π; p]n,1([∂x|...|∂x]
)= 0 + [π, p]
([∂x|...|∂x]
)= ±∂x · p([∂x|...|∂x]) +±p([∂x|...|∂x]) · ∂x
and this is never of the form d · ηj+1 for any p : BA! → A!. It follows that [ad(∂xη
j)] 6= 0 in
Chapter 4. The Characteristic Action of Hochschild Cohomology 100
H∗(hoDer(A!, A!)), and so
Z∞(ExtA(k, k)) =
k[η] if char(k) - d
k[∂x, η] if char(k) | d.
Note that for d = pr and char k = p surjectivity is expected since k[x]/(xd) is isomorphic to the Hopf
algebra kCd, where Cd is the cyclic group or order d.
Example 12. Let A = kQ/(Q3), where Q is the quiver
1
0 2
ba
c
bound by relations Q3 = cba, acb, bac. The 3-homogeneous dual is simply A⊥ = kQop. Using theorem
39, as an algebra A! = ExtA(k,k) is given by kQ′/I where Q′ is
1
0 2
∂a
η1
∂cη0
∂b
η2
where ∂a := Σ−1a∗, η0 = Σ
−2(a∗b∗c∗) comes from the 3-cycle in A⊥, and similarly for the rest. The
relations generating I are given by
∂a∂b = 0, ∂b∂c = 0, ∂c∂a = 0,
and
η0∂a = ∂aη1, η1∂b = ∂bη2, η2∂c = ∂cη0,
all inherited from A⊥. The degree 2 class η := η0 + η1 + η2 is then central and generates a polynomial
subalgebra k[η] ⊆ Z(ExtA(k,k)). The only higher products are given by
m3(∂a ⊗ ∂b ⊗ ∂c) = η0
and its cyclic variants, all extended to be multi-linear over k[η].
The lack of odd oriented cycle shows that η generates Z(ExtA(k,k)). Since |η| is even theorem 40
Chapter 4. The Characteristic Action of Hochschild Cohomology 101
shows that it is in the image of χ. It follows that Z∞(ExtA(k,k)) = k[η] and therefore
HH∗(A,A)/(nil) ∼= k[η].
Example 13. Let A = kQ/(Q3) where Q is the quiver
0 1
3 2
a
bd
c
bound by relations Q3 = dcb, cda, dab, abc. The 3-homogeneous dual is simply A⊥ = kQop. The
algebra A! = ExtA(k,k) is given by kQ′/I where Q′ is
0 1
3 2
∂d
ηa
∂a
ηb
∂c
ηd ∂b
ηc
with ∂a = Σ−1a∗, ηa = Σ
−2b∗c∗d∗ as in example 12, and similarly for all the cyclic variations ∂b, ∂c, ∂d, ηb, ηc, ηd.
The relations are given by
∂a∂b = 0, ∂b∂c = 0, ∂c∂d = 0, ∂d∂a,
and
∂aηa = ηd∂d, ∂dηd = ηc∂c, ∂cηc = ηb∂b, ∂bηb = ηa∂a,
all inherited from A⊥ as in example 12.
Let γ0 := ∂aηa = ηd∂d be the simple cycle at 0, and let η0 := ηdηcηbηa be the long cycle at 0. We
define γ1, γ2, γ3, η1, η2, η3 symetrically. Then let γ =∑3i=0 γi and η =
∑3i=0 ηi , noting that |γ| = 3 and
|η| = 8. Tedious but straightforward calculations in A⊥ show that γ, η generate Z(ExtA(k,k)) ∼= k[γ, η].
Since |η| is even, we know that η is in the image of χ, with only classes of the form ηrγ left to determine.
To compute the higher products we change notation slightly, denoting by ∂j←i the degree 1 arrow
from i to j. We also denote by ηrj←i the monomial path in η0, η1, η2, η3 going from i to j of length r.
Theorem 39 shows that all higher products are sums of products of the form
m3(ηro←n ∂n←m ⊗ ηsm←l ∂l←k ⊗ ηtk←j ∂j←i) = ηr+s+t+1o←i
It follows that ad(ηnγ)(ηsm←l∂l←k ⊗ ηtk←j∂j←i) = 3 · η4n+s+t+1m←i . This is 0 if and only if char(k) | 3, and
when char(k) - 3 similar calculations to example 11 show that [ad(ηnγ)] 6= 0 in H∗(hoDer(A!, A!)). We
Chapter 4. The Characteristic Action of Hochschild Cohomology 102
have shown
Z∞ExtA(k,k) =
k[η] if char(k) 6= 3,
k[γ, η] if char(k) = 3.
Bibliography
[1] J. F. Adams, On the cobar construction, Proc. Nat. Acad. Sci. U.S.A. 42 (1956), 409–412. MR
0079266 15
[2] M. Andre, L’algebre de Lie d’un anneau local, Symposia Mathematica, Vol. IV (INDAM, Rome,
1968/69), Academic Press, London, 1970, pp. 337–375. MR 0276302 77
[3] David J. Anick, A counterexample to a conjecture of Serre, Ann. of Math. (2) 115 (1982), no. 1,
1–33. MR 644015 60
[4] , Hopf algebras up to homotopy, J. Amer. Math. Soc. 2 (1989), no. 3, 417–453. MR 991015
4
[5] E. F. Assmus, Jr., On the homology of local rings, Illinois J. Math. 3 (1959), 187–199. MR 0103907
48
[6] Luchezar Avramov and Stephen Halperin, Through the looking glass: a dictionary between rational
homotopy theory and local algebra, Algebra, algebraic topology and their interactions (Stockholm,
1983), Lecture Notes in Math., vol. 1183, Springer, Berlin, 1986, pp. 1–27. MR 846435 2, 53, 57
[7] Luchezar L. Avramov, Homology of local flat extensions and complete intersection defects, Math.
Ann. 228 (1977), no. 1, 27–37. MR 0485836 58, 72, 74, 75
[8] , Small homomorphisms of local rings, J. Algebra 50 (1978), no. 2, 400–453. MR 485906 5,
60, 82
[9] , Local algebra and rational homotopy, Algebraic homotopy and local algebra (Luminy,
1982), Asterisque, vol. 113, Soc. Math. France, Paris, 1984, pp. 15–43. MR 749041 2, 3, 42, 46,
47, 50, 51, 53, 57, 77
[10] , Golod homomorphisms, Algebra, algebraic topology and their interactions (Stockholm,
1983), Lecture Notes in Math., vol. 1183, Springer, Berlin, 1986, pp. 59–78. MR 846439 38, 39,
50, 58, 59
[11] , Local rings over which all modules have rational Poincare series, J. Pure Appl. Algebra
91 (1994), no. 1-3, 29–48. MR 1255922 50, 60, 66
[12] , Infinite free resolutions, Six lectures on commutative algebra (Bellaterra, 1996), Progr.
Math., vol. 166, Birkhauser, Basel, 1998, pp. 1–118. MR 1648664 30, 31, 33, 35, 41, 44, 50, 83
103
Bibliography 104
[13] , Locally complete intersection homomorphisms and a conjecture of Quillen on the vanishing
of cotangent homology, Ann. of Math. (2) 150 (1999), no. 2, 455–487. MR 1726700 2, 37, 50, 53,
56, 59, 85
[14] Luchezar L. Avramov and Ragnar-Olaf Buchweitz, Homological algebra modulo a regular sequence
with special attention to codimension two, J. Algebra 230 (2000), no. 1, 24–67. MR 1774757 6
[15] , Support varieties and cohomology over complete intersections, Invent. Math. 142 (2000),
no. 2, 285–318. MR 1794064 2
[16] Luchezar L. Avramov, Hans-Bjø rn Foxby, and Bernd Herzog, Structure of local homomorphisms,
J. Algebra 164 (1994), no. 1, 124–145. MR 1268330 8, 36, 37, 50, 63
[17] Luchezar L. Avramov, Vesselin N. Gasharov, and Irena V. Peeva, Complete intersection dimension,
Inst. Hautes Etudes Sci. Publ. Math. (1997), no. 86, 67–114 (1998). MR 1608565 57
[18] Luchezar L. Avramov, Ines Bonacho Dos Anjos Henriques, and Liana M. Sega, Quasi-complete
intersection homomorphisms, Pure Appl. Math. Q. 9 (2013), no. 4, 579–612. MR 3263969 5, 37,
75, 84, 85
[19] Luchezar L. Avramov and Srikanth Iyengar, Finite generation of Hochschild homology algebras,
Invent. Math. 140 (2000), no. 1, 143–170. MR 1779800 4, 33, 61, 64
[20] , Homological criteria for regular homomorphisms and for locally complete intersection
homomorphisms, Algebra, arithmetic and geometry, Part I, II (Mumbai, 2000), Tata Inst. Fund.
Res. Stud. Math., vol. 16, Tata Inst. Fund. Res., Bombay, 2002, pp. 97–122. MR 1940664 34
[21] , Andre-Quillen homology of algebra retracts, Ann. Sci. Ecole Norm. Sup. (4) 36 (2003),
no. 3, 431–462. MR 1977825 52, 53, 55, 56, 58, 59, 63, 66, 67, 69, 70
[22] Luchezar L. Avramov, Srikanth B. Iyengar, and Liana M. ¸Sega, Free resolutions over short local
rings, J. Lond. Math. Soc. (2) 78 (2008), no. 2, 459–476. MR 2439635 60
[23] Luchezar L. Avramov and Li-Chuan Sun, Cohomology operators defined by a deformation, J.
Algebra 204 (1998), no. 2, 684–710. MR 1624432 57
[24] Vladimir Baranovsky, A universal enveloping for L∞-algebras, Math. Res. Lett. 15 (2008), no. 6,
1073–1089. MR 2470385 95
[25] H. J. Baues and J.-M. Lemaire, Minimal models in homotopy theory, Math. Ann. 225 (1977),
no. 3, 219–242. MR 0431172 22, 96
[26] Alexander Beilinson, Victor Ginzburg, and Wolfgang Soergel, Koszul duality patterns in represen-
tation theory, J. Amer. Math. Soc. 9 (1996), no. 2, 473–527. MR 1322847 18, 20
[27] Roland Berger, Koszulity for nonquadratic algebras, J. Algebra 239 (2001), no. 2, 705–734. MR
1832913 6, 96, 97
[28] Alexander Berglund, Poincare series of monomial rings, J. Algebra 295 (2006), no. 1, 211–230.
MR 2188858 41
Bibliography 105
[29] Alexander Berglund and Kaj Borjeson, Free loop space homology of highly connected manifolds,
Forum Math. 29 (2017), no. 1, 201–228. MR 3592599 24
[30] I. N. Bernsteın, I. M. Gel′fand, and S. I. Gel′fand, Algebraic vector bundles on Pn and problems
of linear algebra, Funktsional. Anal. i Prilozhen. 12 (1978), no. 3, 66–67. MR 509387 1
[31] Amalia Blanco, Javier Majadas, and Antonio G. Rodicio, Projective exterior Koszul homology and
decomposition of the Tor functor, Invent. Math. 123 (1996), no. 1, 123–140. MR 1376249 37, 75,
84
[32] Richard E. Block, Commutative Hopf algebras, Lie coalgebras, and divided powers, J. Algebra 96
(1985), no. 1, 275–306. MR 808852 49
[33] B. Briggs and V. Gelinas, Formulas for A-infinity universal envelopes and ext algebras of commu-
tative rings, In preparation. 35
[34] , The A-infinity Centre of the Yoneda algebra and the characteristic Action of Hochschild
Cohomology on the Derived Category, ArXiv e-prints (2017). 1, 5, 8, 25, 26, 27, 28, 29, 86, 91, 95
[35] Winfried Bruns and Jurgen Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Math-
ematics, vol. 39, Cambridge University Press, Cambridge, 1993. MR 1251956 60
[36] Ragnar-Olaf Buchweitz, Hochschild Cohomology of Koszul algebras, talk at the Conference on
Representation Theory, Canberra, July 2003. 90
[37] Ragnar-Olaf Buchweitz and Hubert Flenner, The Atiyah-Chern character yields the semiregularity
map as well as the infinitesimal Abel-Jacobi map, The arithmetic and geometry of algebraic cycles
(Banff, AB, 1998), CRM Proc. Lecture Notes, vol. 24, Amer. Math. Soc., Providence, RI, 2000,
pp. 33–46. MR 1736875 6
[38] Ragnar-Olaf Buchweitz, Edward L. Green, Nicole Snashall, and Øyvind Solberg, Multiplicative
structures for Koszul algebras, Q. J. Math. 59 (2008), no. 4, 441–454. MR 2461267 6, 89
[39] J. Burke, Higher homotopies and Golod rings, ArXiv e-prints (2015). 40
[40] Aldo Conca, Grobner bases for spaces of quadrics of low codimension, Adv. in Appl. Math. 24
(2000), no. 2, 111–124. MR 1748965 60
[41] Vladimir Dotsenko and Bruno Vallette, Higher Koszul duality for associative algebras, Glasg. Math.
J. 55 (2013), no. A, 55–74. MR 3110804 97
[42] Samuel Eilenberg and Saunders Mac Lane, On the groups H(Π, n). I, Ann. of Math. (2) 58 (1953),
55–106. MR 0056295 15
[43] Yves Felix and Stephen Halperin, Rational L.-S. category and its applications, Publ. U.E.R. Math.
Pures Appl. IRMA 2 (1980), no. 3, exp. no. 5, 84. MR 618096 3, 4, 52, 53
[44] Yves Felix, Stephen Halperin, Carl Jacobsson, Clas Lofwall, and Jean-Claude Thomas, The radical
of the homotopy Lie algebra, Amer. J. Math. 110 (1988), no. 2, 301–322. MR 935009 4, 17, 33,
34, 45, 53, 57, 77
Bibliography 106
[45] Yves Felix, Stephen Halperin, and Jean-Claude Thomas, The homotopy Lie algebra for finite
complexes, Inst. Hautes Etudes Sci. Publ. Math. (1982), no. 56, 179–202 (1983). MR 686046 2
[46] Yves Felix, Stephen Halperin, and Jean-Claude Thomas, Rational homotopy theory, Graduate
Texts in Mathematics, vol. 205, Springer-Verlag, New York, 2001. MR 1802847 52
[47] Yves Felix, Steve Halperin, and Jean-Claude Thomas, Exponential growth and an asymptotic
formula for the ranks of homotopy groups of a finite 1-connected complex, Ann. of Math. (2)
170 (2009), no. 1, 443–464. MR 2521122 50
[48] Yves Felix, Luc Menichi, and Jean-Claude Thomas, Gerstenhaber duality in Hochschild cohomol-
ogy, J. Pure Appl. Algebra 199 (2005), no. 1-3, 43–59. MR 2134291 90, 91
[49] Yves Felix, Jean-Claude Thomas, and Micheline Vigue-Poirrier, Loop homology algebra of a closed
manifold, preprint (2003). 6, 95
[50] Domenico Fiorenza and Marco Manetti, L∞ structures on mapping cones, Algebra Number Theory
1 (2007), no. 3, 301–330. MR 2361936 36, 81
[51] Benoit Fresse, Koszul duality of operads and homology of partition posets, Homotopy theory: rela-
tions with algebraic geometry, group cohomology, and algebraic K-theory, Contemp. Math., vol.
346, Amer. Math. Soc., Providence, RI, 2004, pp. 115–215. MR 2066499 43
[52] Murray Gerstenhaber, The cohomology structure of an associative ring, Ann. of Math. (2) 78
(1963), 267–288. MR 0161898 26, 93
[53] Murray Gerstenhaber and Alexander A. Voronov, Homotopy G-algebras and moduli space operad,
Internat. Math. Res. Notices 3 (1995), 141–153 (electronic). MR 1321701 93
[54] Ezra Getzler and John D. S. Jones, A∞-algebras and the cyclic bar complex, Illinois J. Math. 34
(1990), no. 2, 256–283. MR 1046565 27, 28
[55] Victor Ginzburg and Mikhail Kapranov, Koszul duality for operads, Duke Math. J. 76 (1994),
no. 1, 203–272. MR 1301191 5
[56] Paul G. Goerss, On the Andre-Quillen cohomology of commutative F2-algebras, Asterisque (1990),
no. 186, 169. MR 1089001 77
[57] E. S. Golod, Homologies of some local rings, Dokl. Akad. Nauk SSSR 144 (1962), 479–482. MR
0138667 38, 39
[58] E. L. Green, E. N. Marcos, R. Martınez-Villa, and Pu Zhang, D-Koszul algebras, J. Pure Appl.
Algebra 193 (2004), no. 1-3, 141–162. MR 2076383 97, 99
[59] Edward L. Green, William H. Gustafson, and Dan Zacharia, Artin rings of global dimension two,
J. Algebra 92 (1985), no. 2, 375–379. MR 778455 97
[60] Jorge Alberto Guccione and Juan Jose Guccione, Hochschild homology of complete intersections,
J. Pure Appl. Algebra 74 (1991), no. 2, 159–176. MR 1131470 35
[61] , Hochschild homology of complete intersections and smoothness, J. Pure Appl. Algebra 95
(1994), no. 2, 131–150. MR 1293050 35
Bibliography 107
[62] V. K. A. M. Gugenheim and J. Peter May, On the theory and applications of differential tor-
sion products, American Mathematical Society, Providence, R.I., 1974, Memoirs of the American
Mathematical Society, No. 142. MR 0394720 20
[63] T. H. Gulliksen, A homological characterization of local complete intersections, Compositio Math.
23 (1971), 251–255. MR 0301008 50
[64] Tor H. Gulliksen and Gerson Levin, Homology of local rings, Queen’s Paper in Pure and Applied
Mathematics, No. 20, Queen’s University, Kingston, Ont., 1969. MR 0262227 31, 32, 33, 35, 40,
51, 72, 73
[65] Stephen Halperin, The nonvanishing of the deviations of a local ring, Comment. Math. Helv. 62
(1987), no. 4, 646–653. MR 920063 50
[66] Ji-Wei He and Di-Ming Lu, Higher Koszul algebras and A-infinity algebras, J. Algebra 293 (2005),
no. 2, 335–362. MR 2172343 6, 96, 97, 98
[67] Jurgen Herzog and Srikanth Iyengar, Koszul modules, J. Pure Appl. Algebra 201 (2005), no. 1-3,
154–188. MR 2158753 50
[68] Kathryn Hess, A history of rational homotopy theory, History of topology, North-Holland, Ams-
terdam, 1999, pp. 757–796. MR 1721122 52
[69] Vladimir Hinich, Homological algebra of homotopy algebras, Comm. Algebra 25 (1997), no. 10,
3291–3323. MR 1465117 91, 93
[70] , DG coalgebras as formal stacks, J. Pure Appl. Algebra 162 (2001), no. 2-3, 209–250. MR
1843805 43
[71] Johannes Huebschmann, The homotopy type of FΨq. The complex and symplectic cases, Applica-
tions of algebraic K-theory to algebraic geometry and number theory, Part I, II (Boulder, Colo.,
1983), Contemp. Math., vol. 55, Amer. Math. Soc., Providence, RI, 1986, pp. 487–518. MR 862650
96
[72] Dale Husemoller, John C. Moore, and James Stasheff, Differential homological algebra and homo-
geneous spaces, J. Pure Appl. Algebra 5 (1974), 113–185. MR 0365571 16
[73] T. V. Kadeishvili, The algebraic structure in the homology of an A(∞)-algebra, Soobshch. Akad.
Nauk Gruzin. SSR 108 (1982), no. 2, 249–252 (1983). MR 720689 22, 92
[74] , The structure of the A(∞)-algebra, and the Hochschild and Harrison cohomologies, Trudy
Tbiliss. Mat. Inst. Razmadze Akad. Nauk Gruzin. SSR 91 (1988), 19–27. MR 1029003 90
[75] Bernhard Keller, Deriving DG categories, Ann. Sci. Ecole Norm. Sup. (4) 27 (1994), no. 1, 63–102.
MR 1258406 19, 86
[76] , Introduction to A-infinity algebras and modules, Homology Homotopy Appl. 3 (2001),
no. 1, 1–35. MR 1854636 20
[77] , A-infinity algebras in representation theory, Representations of algebra. Vol. I, II, Beijing
Norm. Univ. Press, Beijing, 2002, pp. 74–86. MR 2067371 20, 23
Bibliography 108
[78] , Addendum to: “Introduction to A-infinity algebras and modules” [Homology Homotopy
Appl. 3 (2001), no. 1, 1–35; MR1854636 (2004a:18008a)], Homology Homotopy Appl. 4 (2002),
no. 1, 25–28. MR 1905779 20
[79] , Derived invariance of higher structures on the Hochschild complex, preprint (2003). 12,
14, 28, 30, 88, 90
[80] , A-infinity algebras, modules and functor categories, Trends in representation theory of
algebras and related topics, Contemp. Math., vol. 406, Amer. Math. Soc., Providence, RI, 2006,
pp. 67–93. MR 2258042 13, 29
[81] , On differential graded categories, International Congress of Mathematicians. Vol. II, Eur.
Math. Soc., Zurich, 2006, pp. 151–190. MR 2275593 86
[82] Maxim Kontsevich and Yan Soibelman, Homological mirror symmetry and torus fibrations, Sym-
plectic geometry and mirror symmetry (Seoul, 2000), World Sci. Publ., River Edge, NJ, 2001,
pp. 203–263. MR 1882331 96
[83] Henning Krause and Yu Ye, On the centre of a triangulated category, Proc. Edinb. Math. Soc. (2)
54 (2011), no. 2, 443–466. MR 2794666 12
[84] Tom Lada and Martin Markl, Strongly homotopy Lie algebras, Comm. Algebra 23 (1995), no. 6,
2147–2161. MR 1327129 41, 94
[85] Kenji Lefevre-Hasegawa, Sur les A-infini categories, Ph.D. thesis, Universite Denis Diderot Paris
7, November 2003. 16, 17, 18, 24, 25, 26
[86] Jean-Michel Lemaire, Algebres connexes et homologie des espaces de lacets, Lecture Notes in Math-
ematics, Vol. 422, Springer-Verlag, Berlin-New York, 1974. MR 0370566 52, 60, 66
[87] Gerson Levin, Local rings and Golod homomorphisms, J. Algebra 37 (1975), no. 2, 266–289. MR
0429868 39, 48
[88] , Large homomorphisms of local rings, Math. Scand. 46 (1980), no. 2, 209–215. MR 591601
66, 75
[89] Gerson Louis Levin, HOMOLOGY OF LOCAL RINGS, ProQuest LLC, Ann Arbor, MI, 1965,
Thesis (Ph.D.)–The University of Chicago. MR 2611555 47
[90] Jean-Louis Loday and Bruno Vallette, Algebraic operads, Grundlehren der Mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences], vol. 346, Springer, Heidelberg,
2012. MR 2954392 10, 14, 15, 16, 18, 20, 21, 22, 24, 41, 42, 43, 93, 95
[91] Wendy Lowen, Obstruction theory for objects in abelian and derived categories, Comm. Algebra
33 (2005), no. 9, 3195–3223. MR 2175388 93
[92] Di Ming Lu, John H. Palmieri, Quan Shui Wu, and James J. Zhang, Koszul equivalences in A∞-
algebras, New York J. Math. 14 (2008), 325–378. MR 2430869 18
[93] Lazar Aronovitch Lusternik and L. Schnirelmann, Methodes topologiques dans les problmes varia-
tionnels., Actualites scientifiques et industrielles 118 (1934). 52
Bibliography 109
[94] Dag Madsen, Homological aspects in representation theory, Ph.D. thesis, Trondheim, Norway,
March 2002. 99
[95] J. P. May, The additivity of traces in triangulated categories, Adv. Math. 163 (2001), no. 1, 34–73.
MR 1867203 78
[96] James E. McClure and Jeffrey H. Smith, A solution of Deligne’s Hochschild cohomology conjecture,
Recent progress in homotopy theory (Baltimore, MD, 2000), Contemp. Math., vol. 293, Amer.
Math. Soc., Providence, RI, 2002, pp. 153–193. MR 1890736 28, 93
[97] , Multivariable cochain operations and little n-cubes, J. Amer. Math. Soc. 16 (2003), no. 3,
681–704 (electronic). MR 1969208 93
[98] S. A. Merkulov, Strong homotopy algebras of a Kahler manifold, Internat. Math. Res. Notices
(1999), no. 3, 153–164. MR 1672242 96
[99] John W. Milnor and John C. Moore, On the structure of Hopf algebras, Ann. of Math. (2) 81
(1965), 211–264. MR 0174052 47
[100] Barry Mitchell, Rings with several objects, Advances in Math. 8 (1972), 1–161. MR 0294454 29
[101] Saeed Nasseh and Sean Sather-Wagstaff, Cohen factorizations: weak functoriality and applications,
J. Pure Appl. Algebra 219 (2015), no. 3, 622–645. MR 3279378 51
[102] Cris Negron, Alternate Approaches to the Cup Product and Gerstenhaber Bracket on Hochschild
Cohomology, ProQuest LLC, Ann Arbor, MI, 2015, Thesis (Ph.D.)–University of Washington. MR
3438933 24
[103] Leonid Positselski, Two kinds of derived categories, Koszul duality, and comodule-contramodule
correspondence, Mem. Amer. Math. Soc. 212 (2011), no. 996, vi+133. MR 2830562 21
[104] Stewart B. Priddy, Koszul resolutions, Trans. Amer. Math. Soc. 152 (1970), 39–60. MR 0265437
18, 19, 20
[105] Daniel Quillen, Rational homotopy theory, Ann. of Math. (2) 90 (1969), 205–295. MR 0258031 3,
42, 43
[106] , On the (co-) homology of commutative rings, Applications of Categorical Algebra (Proc.
Sympos. Pure Math., Vol. XVII, New York, 1968), Amer. Math. Soc., Providence, R.I., 1970,
pp. 65–87. MR 0257068 35, 59, 77, 85
[107] , Cyclic cohomology and algebra extensions, K-Theory 3 (1989), no. 3, 205–246. MR
1040400 10
[108] Michael Schlessinger and James Stasheff, The Lie algebra structure of tangent cohomology and
deformation theory, J. Pure Appl. Algebra 38 (1985), no. 2-3, 313–322. MR 814187 41, 91
[109] Colette Schoeller, Homologie des anneaux locaux noetheriens, C. R. Acad. Sci. Paris Ser. A-B 265
(1967), A768–A771. MR 0224682 40, 47, 48
[110] Shoham Shamir, A spectral sequence for the Hochschild cohomology of a coconnective DGA, Math.
Scand. 112 (2013), no. 2, 182–215. MR 3073454 29
Bibliography 110
[111] Gunnar Sjodin, Hopf algebras and derivations, J. Algebra 64 (1980), no. 1, 218–229. MR 575792
47
[112] Nicole Snashall and Øyvind Solberg, Support varieties and Hochschild cohomology rings, Proc.
London Math. Soc. (3) 88 (2004), no. 3, 705–732. MR 2044054 6
[113] James Dillon Stasheff, Homotopy associativity of H-spaces. I, II, Trans. Amer. Math. Soc. 108
(1963), 275-292; ibid. 108 (1963), 293–312. MR 0158400 20
[114] Jim Stasheff, The intrinsic bracket on the deformation complex of an associative algebra, J. Pure
Appl. Algebra 89 (1993), no. 1-2, 231–235. MR 1239562 27
[115] Dmitry E. Tamarkin, Operadic proof of M. Kontsevich’s formality theorem, ProQuest LLC, Ann
Arbor, MI, 1999, Thesis (Ph.D.)–The Pennsylvania State University. MR 2699544 28
[116] P. Tamaroff, Minimal models for monomial algebras, ArXiv e-prints (2018). 7, 96
[117] John Tate, Homology of Noetherian rings and local rings, Illinois J. Math. 1 (1957), 14–27. MR
0086072 33
[118] Bertrand Toen, The homotopy theory of dg-categories and derived Morita theory, Invent. Math.
167 (2007), no. 3, 615–667. MR 2276263 86
[119] Alexander A. Voronov, Homotopy Gerstenhaber algebras, Conference Moshe Flato 1999, Vol. II
(Dijon), Math. Phys. Stud., vol. 22, Kluwer Acad. Publ., Dordrecht, 2000, pp. 307–331. MR
1805923 28
[120] Charles A. Weibel, An introduction to homological algebra, Cambridge Studies in Advanced Math-
ematics, vol. 38, Cambridge University Press, Cambridge, 1994. MR 1269324 10
[121] Hartmut Wiebe, uber homologische Invarianten lokaler Ringe, Math. Ann. 179 (1969), 257–274.
MR 0255531 59
[122] Klaus Wolffhardt, The Hochschild homology of complete intersections, Trans. Amer. Math. Soc.
171 (1972), 51–66. MR 0306192 35
[123] Fei Xu, Hochschild and ordinary cohomology rings of small categories, Adv. Math. 219 (2008),
no. 6, 1872–1893. MR 2455628 6
[124] Justin Young, Brace Bar-Cobar Duality, Ph.D. thesis, Ecole polytechnique federale de Lausanne,
2013. 93
top related